AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
67
Papers today
8h
Update frequency
7
Days of history
PreGS: A Parameter-Transfer-Based Multi-Expert Graph Neural Network for Node Classification
Graph Learning
- Introduces PreGS, a multi-expert GNN framework leveraging parameter transfer.
- Transfers weights from a pretrained GAT to GraphSAGE experts for enhanced structural representation.
- Decouples training by freezing experts, optimizing only the fusion module and MLP classifier.
- Demonstrates competitive performance on eight public graph datasets.
Read more
PreGS: A Parameter-Transfer-Based Multi-Expert Graph Neural Network for Node Classification
Summary
This paper introduces PreGS, a novel multi-expert graph neural network (GNN) framework aimed at improving node classification tasks. Traditional GNNs often rely on a single aggregation mechanism, which may not effectively capture the diverse structural patterns present in various graph datasets. To overcome this limitation, PreGS employs a parameter-transfer approach, where the linear transformation weights from a pretrained multi-head graph attention network (GAT) are transferred to multiple GraphSAGE experts. These experts are then frozen and serve as complementary structural branches. The model integrates raw node features, GAT head representations, and GraphSAGE expert representations into a multilayer perceptron (MLP), which is further fused with the pretrained GAT logits. The authors also propose an enhanced version, PreGSv2, which incorporates source-level weighting and a structural gating mechanism for better adaptive feature integration. Extensive experiments on eight public graph datasets demonstrate that both PreGS and PreGSv2 achieve competitive performance compared to existing GNN baselines. The paper includes various analyses, such as ablation studies and training-time evaluations, validating the framework's effectiveness and stability.
Methodology
The methodology involves pretraining a multi-head GAT and transferring its first-layer attention weights to multiple GraphSAGE experts, which are then frozen during training. The model fuses various representations (raw node features, GAT outputs, and GraphSAGE expert outputs) and utilizes a multilayer perceptron for classification. PreGSv2 enhances this approach with source-level weighting and a structural gating mechanism.
Results
The experiments reveal that PreGS and PreGSv2 outperform several representative GNN baselines across eight public datasets. The results are supported by comprehensive analyses, including ablation studies and evaluations of parameter transfer and training time, confirming the proposed framework's robustness and efficiency.
Implications
The findings suggest that the PreGS framework can be effectively applied to various node classification tasks in diverse domains such as social networks, recommendation systems, and knowledge discovery, enhancing the performance of GNNs by leveraging multi-expert systems.
Concept Drift from a Causal Perspective
Theory
Time Series
- Introduction of a causal taxonomy for concept drift based on Structural Causal Models.
- Development of CaDrift, a synthetic data stream generator for simulating mechanism-level drift events.
- Empirical validation showing that different drift types have unique effects on distributional changes and predictive performance.
- Highlighting the significance of causal reasoning in adaptive learning and model evaluation.
Read more
Concept Drift from a Causal Perspective
Summary
This paper addresses the phenomenon of concept drift in data streams, where changes in the data-generating distribution can adversely affect predictive model performance. Traditional definitions of concept drift focus on changes in the joint distribution P(x, y) without clarifying the underlying causes of these changes. The authors propose a novel causal perspective on concept drift using Structural Causal Models (SCMs), introducing a taxonomy that categorizes drift events based on their causal origins, such as changes in exogenous variables, endogenous mechanisms, confounders, and target-generating processes. They develop a synthetic data stream generator, CaDrift, which simulates controlled mechanism-level drift events, allowing for a more nuanced understanding of how different types of drift affect distributional shifts and predictive behavior. Empirical experiments demonstrate that distinct causal origins of drift lead to varying patterns of marginal and conditional distribution changes, as well as differing impacts on predictive performance. The findings emphasize the importance of considering causal structures in adaptive learning methods and provide a foundation for causally-aware evaluations in non-stationary environments.
Methodology
The authors propose a causal taxonomy of concept drift and develop a synthetic data stream generator called CaDrift. This generator simulates drift events by modeling changes in the structural mechanisms of the data-generating process, allowing for controlled experimentation with different types of drift. Empirical studies are conducted to analyze the effects of these drift types on marginal and conditional distributions and predictive performance.
Results
The experiments reveal that drift events with different causal origins lead to qualitatively different changes in marginal and conditional distributions, as well as varying impacts on predictive performance. Additionally, leveraging the generated data from CaDrift for data augmentation improves downstream performance on real-world datasets.
Implications
The findings suggest that a causal perspective on concept drift can lead to more effective adaptive learning strategies in dynamic environments. By understanding the underlying causal mechanisms, researchers and practitioners can develop more robust models that are better equipped to handle non-stationary data streams.
Targeted Review for AI-Assisted Biodiversity Surveys: Active Continuous-Score Occupancy Modeling
Computer Vision
Audio & Speech
Efficient ML
- ACORN optimizes expert review allocation based on ecological inference goals rather than classifier accuracy.
- The method combines ML predictions with targeted human reviews to enhance occupancy modeling.
- ACORN significantly reduces the number of expert reviews needed while maintaining ecological accuracy.
- The approach is applicable to various population-level inference tasks beyond occupancy modeling.
Read more
Targeted Review for AI-Assisted Biodiversity Surveys: Active Continuous-Score Occupancy Modeling
Summary
This paper addresses the challenge of integrating machine learning (ML) into ecological studies, particularly in the context of occupancy modeling, which estimates species presence based on environmental factors. The authors introduce a novel method called Active Continuous-Score Occupancy Modeling (ACORN), which optimally allocates expert review efforts to maximize ecological inference while minimizing the number of required human reviews. ACORN leverages ML predictions to inform which samples should be reviewed by experts, focusing on those that provide the most valuable information for ecological analysis. The methodology is benchmarked using two datasets from camera traps and bioacoustic sensors, demonstrating that ACORN can achieve ecological conclusions comparable to fully human-labeled data while significantly reducing the review burden. The findings suggest that ML-assisted workflows in ecology should prioritize expert review based on its relevance to ecological inference rather than solely on classifier accuracy, especially when resources for human review are limited.
Methodology
The authors developed ACORN, which integrates noisy ML classifications with a hierarchical occupancy model. It employs Bayesian experimental design to prioritize expert reviews based on their potential to maximize information gain for ecological inference. The method includes a stopping criterion to determine when further expert review is no longer beneficial.
Results
ACORN was evaluated on two ecological datasets and showed that it could recover ecological conclusions nearly as accurately as fully human-labeled data while requiring substantially fewer expert reviews. The method outperformed traditional review policies across multiple metrics related to ecological inference.
Implications
The findings suggest that ACORN can facilitate large-scale biodiversity monitoring by optimizing the use of expert resources, thereby enabling more efficient ecological research and conservation efforts. This approach can be adapted to other ecological modeling tasks that rely on ML-generated data.
In-Context Guidance: Learning Inter-Task Synergies via Numerical Foundational Models for Few-Shot Multitask Optimization
Optimization
Robotics
- Introduction of ICG-MTO framework to improve inter-task coupling estimation under limited evaluation budgets.
- Utilization of a frozen numerical foundational model for auxiliary guidance without replacing the underlying multitask optimizer.
- Development of an in-context query construction strategy for multitask Bayesian optimization.
- Demonstration of effectiveness through experiments on synthetic benchmarks and real-world applications.
Read more
In-Context Guidance: Learning Inter-Task Synergies via Numerical Foundational Models for Few-Shot Multitask Optimization
Summary
This paper presents In-Context Guidance Multitask Optimization (ICG-MTO), a novel framework designed to enhance inter-task coupling estimation in few-shot multitask optimization scenarios. Traditional methods often struggle with limited evaluation budgets, leading to inaccurate inter-task relationship estimations and negative transfer. ICG-MTO leverages a frozen numerical foundational model to provide auxiliary guidance through in-context learning, which is particularly beneficial in data-scarce environments. The framework operates in three stages: constructing an algorithm-specific in-context query from evaluated solutions, using the foundational model to infer a guidance signal that characterizes predictive relationships among tasks, and translating this signal into algorithm-specific guidance for maximum-a-posteriori coupling estimation. This approach allows for regularization during the early stages of optimization and gradually transitions to conventional data-driven behavior as more task-specific observations are collected. The authors instantiate ICG-MTO in multitask Bayesian optimization (ICG-MTBO) and in MFEA-II, demonstrating its effectiveness through experiments on synthetic benchmarks and a real-world robot arm control problem.
Methodology
The ICG-MTO framework consists of three main stages: (1) constructing an in-context query from evaluated solutions, (2) inferring a guidance signal using a frozen numerical foundational model, and (3) translating this signal into algorithm-specific guidance for inter-task coupling estimation. The framework is instantiated in two multitask optimizers: ICG-MTBO and ICG-MFEA-II.
Results
Experiments conducted on synthetic benchmarks and a real-world robot arm control problem showed that ICG-MTO effectively improves inter-task coupling estimation, leading to better optimization performance in few-shot scenarios compared to traditional methods.
Implications
The proposed framework has significant implications for multitask optimization in scenarios with limited data, potentially enhancing performance in various applications such as robotics, where efficient optimization is crucial.
Rewired or Gated? How Instruction Tuning Shapes Knowledge-Conflict Circuits in LLMs
NLP
Large Language Models
Interpretability
- Instruction tuning preserves the conflict circuit by reweighting existing attention heads rather than replacing them.
- Tuning shifts model behavior towards a reliance on parametric memory, affecting how models handle conflicting information.
- The robustness gained from instruction tuning is context-dependent, diminishing with coherent framing of false information.
- Interpretability tools for base models can be effectively transferred to instruct models due to the preservation of the underlying circuit.
Read more
Rewired or Gated? How Instruction Tuning Shapes Knowledge-Conflict Circuits in LLMs
Summary
This paper investigates the impact of instruction tuning on the conflict-resolution circuits of large language models (LLMs). It addresses whether instruction tuning rewires the underlying circuits or merely gates/reweights existing components. The authors conduct a mechanistic comparison of conflict-resolution circuits across three model families: Llama-3.2-3B, Qwen-2.5-3B, and Gemma-3-4B. Utilizing five independent methods, including node and edge attribution and causal ablation, the study finds that instruction tuning primarily reweights the same attention heads rather than replacing them, indicating a preservation of the conflict circuit. The tuning shifts model behavior towards a reliance on parametric memory, leading instruct models to reject terse counterfactual contexts more than base models. However, this skepticism diminishes when false claims are presented in a coherent manner. The findings suggest that interpretability and control tools developed for base models can be directly applied to their instruct variants, highlighting the implications for safety and reliability in model deployment.
Methodology
The authors employed five independent methods for analysis: node and edge attribution, superposition role analysis, causal ablation, and path patching. These methods were used to assess the structure and behavior of the conflict-resolution circuits in the models under study.
Results
The analysis revealed that instruction tuning primarily involves gating rather than rewiring, with a high overlap of nodes (0.60β0.82) across models. The tuning led to a significant reduction in contextual reliance, with instruct models showing a drop from 0.65β0.87 in base models to 0.22β0.37 under terse substitution conflict. The conflict heads remained localized in the upper layers of the network, indicating that the tuning process does not alter their positions.
Implications
The findings suggest that tools for interpreting and controlling LLMs can be directly applied to instruct models, enhancing their usability in practical applications. Additionally, understanding how instruction tuning affects conflict resolution can inform the design of safer and more reliable LLMs, particularly in contexts where conflicting information is prevalent.
MAGIC: Mixed-Granularity Agent Graphs via Incremental Construction with Dense-Reward Reinforcement Learning
Reinforcement Learning
Large Language Models
Graph Learning
- MAGIC allows for mixed-granularity selection in agent graph construction, enhancing flexibility in multi-agent systems.
- The framework utilizes dense-reward reinforcement learning to optimize the construction policy with intermediate feedback.
- MAGIC outperforms existing methods across multiple benchmarks, indicating its superior performance and efficiency.
- The approach addresses the limitations of fixed granularity in existing topology generation methods.
Read more
MAGIC: Mixed-Granularity Agent Graphs via Incremental Construction with Dense-Reward Reinforcement Learning
Summary
The paper introduces MAGIC, a novel dense-reward reinforcement learning framework designed for generating mixed-granularity agent graphs in multi-agent systems (MAS) that utilize large language models (LLMs). The authors argue that existing methods for constructing collaboration topologies in MAS either fix the granularity at the agent or group level, which fails to accommodate the varying collaboration needs of different subtasks. MAGIC addresses this limitation by allowing local selection of granularity for each functional role, enabling a combination of fine-grained control and reusable collaboration patterns. The framework incrementally constructs an agent graph by selecting functional roles, instantiating them as either single agents or groups, and connecting them to existing units. The authors employ potential-based reward shaping to provide intermediate feedback during the construction process, optimizing the policy based on returns from trajectories sampled under the current policy. The results demonstrate that MAGIC outperforms state-of-the-art baselines across eight benchmarks and exhibits strong inference efficiency, highlighting its effectiveness in adapting collaboration structures to task-specific requirements.
Methodology
MAGIC employs a dense-reward reinforcement learning approach to incrementally construct mixed-granularity agent graphs. At each step, it selects a functional role, determines whether to instantiate it as a single agent or a reusable group, and connects it to existing units. The framework uses potential-based reward shaping to provide intermediate feedback based on probe utility and structural complexity, optimizing the construction policy using returns from trajectories sampled under the current policy.
Results
MAGIC demonstrated superior performance compared to state-of-the-art baselines across eight benchmarks, showcasing its effectiveness in generating task-specific agent graphs. Additionally, the framework exhibited strong inference efficiency, indicating its practical applicability in real-world scenarios.
Implications
The findings suggest that MAGIC can significantly enhance the efficiency and effectiveness of multi-agent systems in various applications, particularly those involving complex tasks that require adaptive collaboration structures. This could lead to advancements in fields such as automated programming, testing, and other domains where LLMs are employed.
Mitigating Sequential Reappearance in Diffusion Data-Point Unlearning
Generative Models
Optimization
Theory
- Introduction of a target-level evaluation protocol for sequential diffusion unlearning.
- Identification of sequential reappearance as a failure mode in diffusion unlearning.
- Discovery of data-space recovery sharpness, linking local loss geometry to reappearance risk.
- Development of LASTING, a method that improves forgetting persistence while preserving generation utility.
Read more
Mitigating Sequential Reappearance in Diffusion Data-Point Unlearning
Summary
This paper addresses the challenge of diffusion data-point unlearning, specifically focusing on the issue of sequential reappearance, where previously forgotten instances re-emerge in the model after subsequent deletion requests. The authors introduce a target-level evaluation protocol that assesses whether a target is forgotten immediately and remains forgotten throughout a sequence of deletions. They identify a phenomenon termed 'data-space recovery sharpness,' which characterizes the local loss geometry of targets that reappear versus those that remain forgotten. To combat this issue, the authors propose a novel method called LASTING (Low-loss Adaptive Search and Targeted Intervention in Neighborhood Geometry), which iteratively identifies and suppresses the most recoverable representations around each target during the unlearning process. Experimental results demonstrate that LASTING significantly enhances forgetting persistence across long deletion sequences while maintaining the generation quality of the model, thus emphasizing the need for robust evaluation of unlearning methods beyond immediate deletion effectiveness.
Methodology
The authors developed a target-level evaluation protocol to track the forgetting status of targets over sequential deletions. They characterized the local loss geometry of targets post-deletion to identify vulnerabilities associated with reappearance. LASTING was proposed as a method that adaptively searches for and suppresses recoverable representations around each target during the unlearning process, iteratively refining its approach as the model updates.
Results
Experiments showed that LASTING significantly improved the persistence of forgetting across long sequences of deletions compared to existing methods, while also preserving the generation quality of the diffusion model. The results indicated that targets that reappeared had sharper local loss geometries, confirming the association between recovery sharpness and reappearance risk.
Implications
The findings suggest that effective data-point unlearning in diffusion models requires not only immediate deletion success but also sustained forgetting through subsequent updates. This has implications for privacy and copyright concerns in machine learning applications, particularly in scenarios involving sensitive data.
Federating Quantum and Classical Computing: A Privacy-Preserving Hybrid Approach
Federated Learning
- The paper introduces a hybrid quantum-classical model that leverages federated learning for privacy preservation.
- The SBVFL protocol is utilized to avoid centralizing raw data and reduce communication overhead.
- Simulation results show a significant accuracy improvement while maintaining parameter efficiency.
- The study establishes conditions under which quantum models can outperform classical models in a federated setting.
Read more
Federating Quantum and Classical Computing: A Privacy-Preserving Hybrid Approach
Summary
This paper explores the integration of quantum machine learning (QML) with classical computing through a federated learning (FL) framework, specifically focusing on privacy-preserving methods. The authors propose a hybrid model where a quantum active party collaborates with a classical passive party using Sherpa.aiβs Blind Vertical FL (SBVFL) protocol. This approach allows for the training of models without centralizing raw data, addressing privacy concerns in multi-party settings. The study introduces the split multiplicative periodic parity (SMPP) benchmark to evaluate the effectiveness of the proposed method. Simulations demonstrate that the SBVFL protocol significantly improves model accuracy from 0.7227 to 0.8757 compared to local training, nearing the accuracy of non-private centralized models while utilizing fewer trainable parameters than classical alternatives. The findings indicate that FL can facilitate effective quantum-classical collaboration while maintaining data privacy.
Methodology
The authors employed a hybrid quantum-classical architecture, integrating a parameterized quantum circuit within a classical model. They utilized the SBVFL protocol to facilitate training without exposing raw data. The effectiveness of the approach was evaluated using the SMPP benchmark through simulations that compared federated learning outcomes against local training results.
Results
The simulations revealed that the SBVFL protocol improved model accuracy from 0.7227 to 0.8757, closely approaching the accuracy of non-private centralized models. The hybrid model achieved this with significantly fewer trainable parameters compared to classical neural networks and random forest models, demonstrating the potential for efficient training in a privacy-preserving context.
Implications
The findings suggest that federated learning can effectively enable collaboration between quantum and classical computing systems while preserving privacy. This has potential applications in various fields, including finance, healthcare, and any domain requiring secure multi-party data analysis.
Terminal Shrinkage Averaging Reveals a Schedule-Estimator Interaction in LLM Pretraining
Large Language Models
Optimization
Theory
- Introduction of Terminal Shrinkage Averaging (TSA) as a new estimator for LLM pretraining.
- Demonstration of the interaction between learning-rate schedules and output estimators.
- Controlled experiments show that TSA improves validation quality over the raw final iterate.
- Findings suggest that a more active terminal learning-rate schedule is beneficial when using TSA.
Read more
Terminal Shrinkage Averaging Reveals a Schedule-Estimator Interaction in LLM Pretraining
Summary
This paper addresses the interaction between learning-rate schedules and output estimators in the pretraining of large language models (LLMs). Traditionally, LLM pretraining utilizes the raw final iterate as the deployed model, which couples the learning-rate schedule with the estimator choice. The authors propose Terminal Shrinkage Averaging (TSA), a method that interpolates between the raw final iterate and the average of recent checkpoints, allowing for a balance between optimization progress and reduction of variation in the final model. The study analyzes how TSA influences the preferred terminal learning-rate schedule through a local quadratic approximation and tests this interaction using controlled NanoChat experiments. The results indicate that TSA allows for a more active terminal trajectory, improving validation quality in depth-22 NanoChat models and demonstrating faster completion times compared to existing baselines. This work highlights the importance of separating the learning-rate schedule from the output estimator to enhance model performance.
Methodology
The authors developed TSA as an interpolation method between the raw final iterate and the average of recent checkpoints. They conducted controlled experiments using depth-12 NanoChat models to isolate the effects of averaging and terminal learning rates. A local quadratic approximation of the loss was used to derive the bias-variance tradeoff associated with TSA.
Results
The experiments revealed that using TSA improved validation bits per byte (BPB) when the terminal learning-rate floor was increased, contrasting with the performance of the raw final iterate. The findings were replicated under a pure AdamW optimizer, confirming the robustness of the results. Additionally, applying TSA to depth-22 NanoChat models led to improved validation quality and faster completion times compared to public baselines.
Implications
The findings suggest that separating the learning-rate schedule from the output estimator can lead to significant improvements in LLM pretraining efficiency and model performance. This approach may influence future training strategies for large-scale language models and other deep learning applications.
Dissecting Hierarchical Reasoning Models: A Mechanistic Study
Theory
Interpretability
- HRM demonstrates superior performance in reasoning tasks compared to one-pass Transformer models.
- The functional contributions of high- and low-level states in HRM vary significantly across different tasks.
- Linearly decodable features do not necessarily indicate causal relevance in the model's inference process.
- Ablations of features identified by sparse autoencoders yield larger behavioral changes than those from linear probes.
Read more
Dissecting Hierarchical Reasoning Models: A Mechanistic Study
Summary
This paper investigates the Hierarchical Reasoning Model (HRM), a hierarchical Transformer-based model designed for latent-space reasoning, which allows for iterative computation without explicit intermediate verbalization. The authors aim to understand the internal mechanisms of HRM and how it encodes information while performing reasoning tasks such as Sudoku, Maze navigation, and ARC-AGI-2. Through a series of experiments, the authors compare HRM against various Transformer baselines, apply causal interventions, and utilize linear probes and sparse autoencoders (SAEs) for feature analysis. The findings reveal that recurrent models outperform one-pass baselines, and the contributions of high- and low-level states vary across tasks and inference stages. Notably, while task variables can be linearly decoded from HRM's recurrent states, the causal relevance of these features is not supported by directed ablation experiments. The study concludes that HRM operates through constraint-aware iterative refinement, emphasizing the need for enhanced mechanistic interpretability techniques for latent-space reasoning models.
Methodology
The authors conducted a mechanistic study of HRM by comparing it against Transformer baselines with and without recurrent modules. They applied causal interventions on latent representations, performed representational analysis using linear probes and sparse autoencoders, and analyzed the effects of feature ablations on model performance across different reasoning tasks.
Results
The results indicate that HRM iteratively refines a puzzle-specific solution state, with varying contributions from high- and low-level modules depending on the task. While task-relevant information can be decoded from the high-level state, the directed ablation of probe directions did not yield significant differences from random controls. Additionally, the top features identified through SAEs did not show a consistent advantage over random feature subsets, suggesting a distributed computation model in HRM.
Implications
The findings underscore the importance of understanding the internal mechanisms of hierarchical reasoning models, which could lead to improved interpretability and performance in complex reasoning tasks. This research may inform the development of more effective latent-space reasoning models and enhance their applicability in fields requiring sophisticated problem-solving capabilities.
Exactness at Inference: A Representational Criterion for Out-of-Distribution Generalization
Theory
- Exactness at inference is necessary for both in-distribution and out-of-distribution generalization.
- Structural equivalence to the data-generating mechanism is crucial for achieving exactness.
- Current deep learning architectures often fail to generalize due to reliance on fitted approximations.
- The proposed inference-path propagation rule helps in understanding representational limits in hybrid models.
Read more
Exactness at Inference: A Representational Criterion for Out-of-Distribution Generalization
Summary
This paper introduces a novel criterion for out-of-distribution (OOD) generalization, termed 'exactness at inference'. The authors argue that a model can only generalize beyond its training distribution if it computes a representation that is structurally equivalent to the underlying data-generating mechanism, rather than relying on fitted approximations. They demonstrate that structural equivalence is essential for both in-distribution and OOD exactness. The paper discusses Tensor Logic as an example of this independence, where zero-temperature tensor contractions achieve strict mathematical equivalence to discrete logic. The authors highlight that existing deep learning architectures often fail to meet this criterion due to their reliance on approximations that do not hold when faced with novel entities or extrapolation tasks. They propose an inference-path propagation rule for hybrid architectures, which allows the output to inherit representational bounds from fitted estimators along its path. The findings reveal limitations in current architectures and emphasize the need for models that can autonomously induce exact representations. The paper concludes with an example demonstrating that exact models can identify definitively answerable queries, unlike deep network ensembles that exhibit false confidence.
Methodology
The authors develop a theoretical framework to define and evaluate the concept of exactness at inference. They analyze various models, including Tensor Logic and differentiable Inductive Logic Programming (ILP), to illustrate the application of their criterion. The paper also discusses the implications of exactness in the context of hybrid architectures and provides empirical examples to support their claims.
Results
The paper demonstrates that models adhering to the exactness criterion, such as zero-temperature Tensor Logic and differentiable ILP, outperform traditional deep networks in terms of generalization to OOD data. It highlights specific failure modes of deep networks, such as piecewise-affine extrapolation divergence and inability to bind novel entities, which stem from their lack of exact representability.
Implications
The findings suggest that future machine learning architectures should focus on inducing exact representations autonomously rather than relying on approximations. This could lead to more robust models capable of generalizing effectively to unseen data, with potential applications in various domains requiring reliable predictions under uncertainty.
The Probabilistic Structure of Large Language Models
Large Language Models
Generative Models
Theory
- Introduces a unified probabilistic framework for understanding LLMs.
- Frames training as a maximum-likelihood estimation problem using stochastic gradient methods.
- Examines the role of KullbackβLeibler divergence in text generation and its implications.
- Discusses diffusion models and their probabilistic structure in contrast to autoregressive models.
Read more
The Probabilistic Structure of Large Language Models
Summary
This paper provides a comprehensive probabilistic framework for understanding large language models (LLMs), integrating various tools and concepts that are often treated separately in the literature. It describes LLMs as probability measures on sequences of tokens, defined through autoregressive conditional distributions. The training process is framed as a maximum-likelihood estimation problem, solved using stochastic gradient methods, while text generation is conceptualized as simulating a stochastic process. The paper explores the implications of KullbackβLeibler divergence in text generation, particularly regarding phenomena like hallucination and the distinction between statistical plausibility and factual accuracy. Additionally, it discusses diffusion models, emphasizing their probabilistic structure and contrasting them with autoregressive models. The author aims to clarify the probabilistic underpinnings of LLMs, addressing common misconceptions and providing a detailed exposition of the relevant concepts, including the relationship between empirical and theoretical distributions, and the mechanics of word selection during generation.
Methodology
The paper employs a probabilistic approach to describe LLMs, utilizing maximum-likelihood estimation for training and stochastic simulation for text generation. It integrates concepts from probability theory, including autoregressive processes and KullbackβLeibler divergence, to analyze the behavior of LLMs and their generative capabilities.
Results
The paper successfully articulates a clear probabilistic framework for LLMs, highlighting the differences between statistical plausibility and truth in generated text. It also provides insights into the mechanics of text generation and the implications of model training on output quality, particularly regarding hallucination and factual accuracy.
Implications
The findings have significant implications for the development and evaluation of LLMs, emphasizing the need for a deeper understanding of their probabilistic foundations. This could lead to improved training methodologies and more reliable text generation, with applications across various fields such as natural language processing, AI-driven content creation, and automated reasoning.
The Sirens' Song: When Proximal Background Context Overshadows Distant Evidence
NLP
Large Language Models
- Identification of the Proximity Trap, where proximal context overshadows distant evidence.
- Introduction of LYRA, a mechanism that reshapes context retrieval to prioritize task-relevant evidence.
- Demonstration of improved performance in long-context understanding through masking proximal background.
- Development of ProxBench for evaluating distant evidence utilization under proximal interference.
Read more
The Sirens' Song: When Proximal Background Context Overshadows Distant Evidence
Summary
This paper addresses a phenomenon termed the Proximity Trap, where proximal, task-irrelevant context competes with and overshadows distant, relevant evidence in long-context large language models (LLMs). The authors argue that the challenge of utilizing distant evidence is not solely due to its distance but also significantly influenced by the abundance of nearby irrelevant context. To mitigate this issue, they introduce LYRA (Long-context heavY-tailed Relevance Alignment), a t-distributed directional matching mechanism that reshapes the context retrieval distribution, enhancing the model's focus on task-relevant evidence while maintaining positional information. The paper presents extensive experiments across various benchmarks, demonstrating that reducing attention to proximal background can improve long-context understanding. Additionally, the authors introduce ProxBench, a benchmark designed to evaluate distant evidence utilization amidst increasing proximal background interference, further validating their findings.
Methodology
The authors propose LYRA, which employs a t-distributed directional matching mechanism to adjust the attention scores between queries and keys, thereby reducing the influence of irrelevant proximal context. They conduct experiments on LongBench-v2, RULER, and LongBench to assess the effectiveness of LYRA and the impact of masking proximal tokens on model performance.
Results
The experiments show consistent improvements in accuracy across different context lengths and task categories when proximal background tokens are masked. The introduction of LYRA leads to better utilization of distant evidence, supporting the hypothesis that reducing attention to irrelevant proximal context enhances overall model performance.
Implications
The findings suggest that improving long-context LLMs may require a reevaluation of how attention is allocated, particularly in the presence of irrelevant proximal context. This could lead to more effective models in tasks requiring the integration of distant evidence, with potential applications in various NLP tasks such as document summarization, question answering, and information retrieval.
When Recursive Models Finish Computing
Theory
- Extending the computation steps significantly improves the accuracy of recursive models.
- Many incorrect outputs at nominal budgets are due to unfinished computations rather than model failures.
- The concept of trajectory-conditioned anisotropic stability characterizes the stability dynamics of completed computations.
- Different models exhibit varying behaviors in terms of perturbation responses after completion.
Read more
When Recursive Models Finish Computing
Summary
This paper investigates the dynamics of completion in Tiny Recursive Models (TRMs) using attention and MLP architectures, particularly in the context of solving hard Sudoku puzzles. The authors highlight a critical ambiguity in recursive models: an incorrect output at the end of a nominal inference budget may indicate either a failure or an unfinished computation. By extending the recurrence from a nominal 16 steps to 512 steps, the authors demonstrate significant improvements in cumulative exact-solve accuracy, achieving 87.5% for the attention model and 91.9% for the MLP model. The study identifies a phenomenon termed 'trajectory-conditioned anisotropic stability,' where completed states exhibit local contractiveness along the trajectory direction, despite the presence of expanding directions in the local Jacobian. The analysis reveals that many apparent failures within the nominal budget are actually unfinished computations, with over two-thirds of unresolved puzzles being solvable with extended computation. The paper also discusses the differing behaviors of early and late solvers and the implications of these findings for understanding the dynamics of recursive models.
Methodology
The authors conducted experiments on 1,000 hard Sudoku puzzles using attention-based and MLP-based TRMs. They extended the nominal computation steps from 16 to 512 and analyzed the completion dynamics, latent motion, and local stability using local Jacobians and perturbation experiments. The study categorized puzzles into early, late, and persistent solvers based on when they were first solved.
Results
The attention model's cumulative exact-solve accuracy increased from 59.2% to 87.5%, while the MLP model improved from 74.4% to 91.9%. Over 68% of puzzles unresolved at 16 steps were solved with extended computation. The study identified a common stability signature for completed computations, with completed states showing reduced latent motion and directional stability.
Implications
The findings suggest that recursive models can benefit from extended computation to improve accuracy, and they provide insights into the dynamics of model completion. This understanding can inform the design of more effective recursive architectures and enhance their application in complex reasoning tasks.
Brain-Inspired Hierarchical Modularity for General Continual Learning
Robotics
Multimodal
Computer Vision
- Introduction of FlyGCL, a framework for general continual learning inspired by biological systems.
- Hierarchical modularity enables effective separation of conflicting experiences and integration of compatible ones.
- Significant performance improvements in various tasks, particularly in embodied vision-language-action learning.
- The framework leverages pretrained models for stable representation and efficient learning.
Read more
Brain-Inspired Hierarchical Modularity for General Continual Learning
Summary
This paper presents FlyGCL, a brain-inspired framework for general continual learning (GCL) that addresses the challenges of learning from online, uncertain, and evolving data streams. Traditional continual learning methods often operate under simplified conditions with clear task boundaries, which limits their applicability in real-world scenarios. The authors draw inspiration from the Drosophila learning and memory system, proposing a hierarchical modular principle that balances expert specialization and ensemble integration. FlyGCL utilizes pretrained foundation models and incorporates lightweight modular adaptations, including brain-inspired random expansion for expert routing and diversified modular integration across spatial and temporal scales. The framework is evaluated across various tasks, including visual recognition and embodied vision-language-action learning, demonstrating significant improvements in performance, particularly in scenarios without memory replay. The results indicate that hierarchical modularity can effectively organize learning from dynamic experiences, providing a biologically grounded approach to continual learning.
Methodology
The authors propose FlyGCL, which employs a hierarchical modular structure that combines expert specialization and ensemble integration. This is achieved through lightweight adaptations of pretrained models, utilizing brain-inspired random expansion for routing among specialized experts and introducing differentiated adaptive modules for multi-timescale predictions. The methodology emphasizes the organization of learning to accommodate evolving data distributions.
Results
FlyGCL consistently outperforms traditional replay-free continual learning methods across multiple tasks, achieving final average success rates of 83.1%, 86.1%, 94.5%, and 79.1% in various benchmarks. The gains over the strongest baseline range from 49.5 to 58.8 percentage points, particularly highlighting its effectiveness in embodied vision-language-action learning.
Implications
The findings suggest that hierarchical modularity can serve as a foundational principle for developing intelligent systems capable of adapting to dynamic environments. This approach may enhance the robustness and efficiency of AI applications in real-world scenarios, particularly in fields requiring continual learning from diverse and evolving data.
Hierarchical Bayesian optimization of an aircraft-based multi-agent system-of-systems
Optimization
- Introduces a hierarchical Bayesian optimization framework for aircraft-based multi-agent systems.
- Utilizes Gaussian process meta-modeling to handle complex design variables and dependencies.
- Demonstrates improved search efficiency and robustness over conventional optimization methods.
- Applies the framework to a real-world case study in wildfire suppression.
Read more
Hierarchical Bayesian optimization of an aircraft-based multi-agent system-of-systems
Summary
This paper presents a novel hierarchical Bayesian optimization framework designed for optimizing complex aircraft-based multi-agent systems within the context of System-of-Systems (SoS). The authors address the challenges posed by high-fidelity multiphysics simulations, which are computationally expensive and prone to evaluation failures. To mitigate these issues, the proposed framework utilizes Gaussian process meta-modeling to effectively manage discrete architectural choices, conditional dependencies, and heterogeneous design variables. The framework is demonstrated through a case study focused on optimizing an aircraft-based multi-agent system for wildfire suppression, part of the EU-funded COLOSSUS project. The results indicate that the hierarchical optimization approach significantly enhances search efficiency and robustness compared to traditional surrogate-based methods, allowing for effective exploration of large and diverse design spaces with limited simulation resources. The methodology not only advances the state of practice in SoS architecting but also provides insights applicable to various domains, including aviation and emergency response.
Methodology
The methodology involves a hierarchical Bayesian optimization framework that employs Gaussian process meta-modeling to address the complexities of System-of-Systems design. This includes managing discrete architectural choices and conditional dependencies while optimizing heterogeneous design variables. The approach is validated through simulations and case studies, particularly focusing on the optimization of multi-agent systems for specific use cases.
Results
The hierarchical Bayesian optimization framework demonstrated significant improvements in search efficiency and robustness, enabling effective exploration of large design spaces with limited simulation budgets. The case study on wildfire suppression showcased the practical applicability of the framework, highlighting its effectiveness in coordinating heterogeneous aerial platforms.
Implications
The findings suggest that the hierarchical Bayesian optimization framework can be a valuable tool for designing and optimizing complex System-of-Systems architectures in aviation and other fields. Its scalability and efficiency can facilitate the development of innovative solutions for challenges like sustainable mobility and emergency response, ultimately contributing to advancements in system design and operational efficiency.
Slow Decay and Silenced Expression: Iterated Subliminal Trait Transfer in Language-Model Lineages
NLP
Large Language Models
Theory
- Traits can persist through multiple generations of language models, albeit with diminishing behavioral expression.
- The study demonstrates a divergence between internal representations and external expressions of traits in language models.
- Subliminal learning allows for the transmission of traits through filtered data, raising questions about the safety and implications of such processes.
- The research highlights the importance of evaluating both behavioral and internal metrics to understand model capabilities.
Read more
Slow Decay and Silenced Expression: Iterated Subliminal Trait Transfer in Language-Model Lineages
Summary
This paper investigates the persistence of traits in language models trained iteratively on outputs from previous models, termed 'lineages'. The authors explore subliminal learning, where a trait can be transmitted from a teacher model to a student model through filtered data that does not contain the trait's content. They extend previous research by examining whether such traits hold or fade across multiple generations of models. Using three copies of the Qwen2.5-7B-Instruct model, the authors perform ten iterations of training and evaluate the models' outputs through two methods: a keyword screen for behavioral expression of the trait and an activation probe for internal representation. The findings reveal that while the trait persists across generations, its behavioral expression diminishes significantly, indicating that a trait can be internally present but not externally expressed. This divergence between internal and external measures suggests that behavioral evaluations alone may not fully capture the presence of traits in language models.
Methodology
The authors trained three independent lineages of the Qwen2.5-7B-Instruct model over ten iterations, instilling a specific trait (owl preference) and evaluating each generation using a keyword screen and an activation probe. The keyword screen assessed behavioral expression, while the activation probe measured internal representation of the trait.
Results
The study found that the instilled trait persisted across ten generations, with keyword-screen rates dropping from 55.6% after the first step to 21.1% by generation ten. Notably, the base model showed no expression of the trait in 300 completions. Furthermore, when the system prompt was removed, the generation-ten students exhibited zero keyword-screen rates, despite positive activation probe scores, indicating that the trait could be present internally while absent behaviorally.
Implications
The findings suggest that language models can retain traits through iterative training, but the diminishing behavioral expression raises concerns about the reliability of such traits in practical applications. This has implications for the design and evaluation of AI systems, particularly in contexts where trait transmission could influence model behavior and safety.
Label-Efficient Learning for Ground-Based Sky-Image Classification: A Benchmark of Transfer Learning, Active Learning, and Pseudo-Labeling on GCD
Efficient ML
Computer Vision
- Supervised transfer learning significantly enhances label efficiency for cloud classification.
- Active learning and pseudo-labeling provide limited improvements over a strong supervised baseline.
- Accepted pseudo-labels are reliable but biased towards easier cloud types.
- Uncertainty sampling targets challenging samples but yields modest performance gains.
Read more
Label-Efficient Learning for Ground-Based Sky-Image Classification: A Benchmark of Transfer Learning, Active Learning, and Pseudo-Labeling on GCD
Summary
This paper addresses the challenge of accurate ground-based cloud classification, which is crucial for various applications including atmospheric monitoring and solar energy forecasting. The authors benchmark three label-efficient learning strategiesβsupervised transfer learning, uncertainty-based active learning, and high-confidence pseudo-labelingβusing the Ground-based Cloud Dataset (GCD). They employ a frozen ImageNet-pretrained ResNet50 model and evaluate performance across different label budgets ranging from 1% to 100%. The study finds that supervised transfer learning is highly effective, achieving a test accuracy of 0.730 with 40% of labels, closely approaching the full-label performance of 0.735. Active learning and pseudo-labeling show competitive results but do not consistently outperform the supervised baseline. The analysis reveals that accepted pseudo-labels are reliable but biased towards easier classes, while uncertainty sampling targets challenging samples with only modest performance gains. Overall, the findings suggest that transfer learning significantly reduces annotation needs, while simple active and semi-supervised strategies offer limited additional benefits.
Methodology
The study employs a frozen ImageNet-pretrained ResNet50 model to benchmark three learning strategies: supervised transfer learning, uncertainty-based active learning, and high-confidence pseudo-labeling. Experiments are conducted over five random seeds with varying label budgets from 1% to 100% of the training labels, and performance is evaluated using overall and class-balanced metrics.
Results
The results indicate that supervised transfer learning achieves a test accuracy of 0.730 with 40% of labels, closely matching the full-label performance of 0.735. Active learning and pseudo-labeling show competitive performance but do not provide substantial aggregate gains. Pseudo-labels demonstrate high reliability with accuracy ranging from 0.946 to 0.977, while uncertainty sampling focuses on challenging samples but results in only modest improvements.
Implications
The findings suggest that transfer learning can greatly reduce the need for labeled data in cloud classification tasks, which is beneficial for practical applications in atmospheric monitoring and related fields. The study also highlights the need for careful consideration of label-efficient strategies in imbalanced datasets.
Success Leaves Detours: Learning Executable Walkthroughs for Long-Horizon Agents
Reinforcement Learning
Large Language Models
Robotics
- Introduces Trace, a framework for inducing executable walkthroughs from sparse-reward trajectories.
- Combines credit propagation with dependency analysis to filter necessary actions from noisy interaction histories.
- Demonstrates significant performance improvements over eight baselines in long-horizon tasks.
- Walkthroughs generated by Trace support cross-episode reuse and programmatic verification.
Read more
Success Leaves Detours: Learning Executable Walkthroughs for Long-Horizon Agents
Summary
This paper addresses the challenge of improving test-time self-evolving agents by transforming sparse-reward trajectories into executable walkthroughs. The authors propose a novel framework called Trace, which focuses on extracting compact, state-conditioned, and verifiable procedures from noisy interaction histories. The key insight is that while delayed credit can indicate an action's historical relevance, it does not confirm its necessity for subsequent actions. Trace employs a three-stage process: it identifies progress anchors from rewards and state changes, propagates credit backward to highlight valuable transitions, and estimates action prerequisites based on cross-episode evidence. The framework effectively filters out irrelevant detours and loops, resulting in executable walkthroughs that encode entry conditions, state-action-effect steps, and dependencies. Experimental results demonstrate that Trace significantly outperforms existing baselines in various long-horizon tasks, showcasing its ability to enhance interaction through state-conditioned procedures rather than relying on complete trajectories or abstract summaries.
Methodology
The methodology involves a three-stage pipeline: (1) canonicalizing observations and detecting progress anchors, (2) propagating credit backward to identify high-value transitions, and (3) estimating action prerequisites and filtering out irrelevant steps through backward dependency slicing.
Results
Trace outperformed eight test-time learning and memory baselines, achieving improvements of 30.0% in average AUC and 40.5% in Final-3 metrics compared to the strongest baseline, while also using fewer inference tokens. Ablation studies confirmed the contributions of various components of the framework.
Implications
The findings suggest that long-horizon agents can benefit from structured, state-conditioned procedures, which may lead to more efficient learning and execution in complex environments. This approach could be applied in various domains such as web navigation, software operation, and interactive problem-solving.
LoRA-generating hypernetworks for efficient on-device LLM generative personalization
NLP
Large Language Models
Efficient ML
- Introduces LoRA-generating hypernetworks for on-device LLM personalization.
- Combines benefits of ICL and PEFT while avoiding their drawbacks.
- Demonstrates efficiency in computational cost and latency for mobile devices.
- Focuses on long-form text generation tasks, showing competitive performance.
Read more
LoRA-generating hypernetworks for efficient on-device LLM generative personalization
Summary
This paper introduces a novel method for personalizing on-device large language models (LLMs) using LoRA-generating hypernetworks. The authors highlight the challenges of limited computational resources on mobile devices, which restrict model scale and quality. By training a hypernetwork to map user context tokens to a low-rank adaptation (LoRA), the method allows for efficient personalization without the computational burden of traditional parameter-efficient fine-tuning (PEFT) or the drawbacks of in-context learning (ICL). The hypernetwork generates personalized LoRAs entirely on-device, requiring only forward passes through neural networks, thus maintaining low latency and minimal storage requirements. The paper presents comparative evaluations against ICL and PEFT, particularly focusing on long-form text generation tasks, demonstrating that the proposed method achieves comparable or superior quality in personalized responses while being more efficient in terms of computational cost and latency.
Methodology
The methodology involves training a hypernetwork that takes user context tokens as input and outputs a personalized LoRA. This process is conducted entirely on-device, allowing for efficient personalization without the need for gradient descent or increasing input sequence length. The approach is structured into three phases: common artifact creation, personal artifact creation, and personalized inference.
Results
The experiments conducted on various personalization datasets indicate that the hypernetwork-generated LoRAs provide responses that are at least as good, if not better, in quality compared to traditional methods like ICL and PEFT. The results highlight the method's effectiveness in maintaining low computational costs and latency while delivering high-quality personalized outputs.
Implications
The findings suggest that LoRA-generating hypernetworks could significantly enhance the personalization of on-device LLMs, making them more effective for individual users while adhering to the constraints of mobile environments. This could lead to improved user experiences in applications requiring generative AI features on mobile devices.
Lightweight Ranking Heads: Accelerating Multi-Task Experimentation in Production Recommender Systems
Efficient ML
Multimodal
Optimization
- Introduction of the Lightweight Ranking Heads framework for dynamic task addition.
- Elimination of cold-starting and retraining requirements for existing models.
- Utilization of stop-gradients to maintain training stability.
- Centralized configuration for simultaneous task integration across models.
Read more
Lightweight Ranking Heads: Accelerating Multi-Task Experimentation in Production Recommender Systems
Summary
The paper introduces the Lightweight Ranking Heads (Light Heads) framework, designed to enhance the experimentation velocity in production-scale recommender systems. Traditional multi-task ranking models face challenges when integrating new prediction tasks, often leading to bottlenecks due to task conflicts and the need for extensive retraining of backbone models. The Light Heads framework allows for the dynamic addition of new ranking tasks without requiring cold-starting or retraining existing models. By employing techniques such as stop-gradients and stateless daily training, the framework isolates new tasks to prevent adverse interactions with existing ones. A centralized configuration enables simultaneous addition of Light Heads across multiple models, significantly accelerating the training data generation and co-training processes. The implementation of this framework at YouTube has reduced the iteration cycle for multi-task experimentation from weeks to days, facilitating rapid A/B testing and deployment of new ranking tasks that deliver measurable production value.
Methodology
The methodology involves creating lightweight ranking heads that can be injected into existing multi-task models without retraining. The framework employs stop-gradients and stateless training to ensure stability and consistency. A centralized configuration allows for the simultaneous deployment of new heads across multiple models, facilitating faster experimentation.
Results
The implementation of Light Heads at YouTube resulted in a drastic reduction in the iteration cycle for multi-task experimentation, decreasing it from several weeks to just days. This enabled rapid A/B testing and deployment of new ranking tasks, leading to measurable improvements in production outcomes.
Implications
The Lightweight Ranking Heads framework has the potential to revolutionize the way multi-task models are developed and experimented with in production environments, allowing for quicker iterations and more efficient resource utilization. This could lead to enhanced performance in recommender systems and other applications requiring multi-task learning.
Q-DEQ: Discrete Solving and Quantization for Deep Equilibrium Models in Time Series Forecasting under Edge Deployment Coding Constraints
Time Series
Optimization
Efficient ML
- Q-DEQ introduces a local discrete solver for DEQ forward computation, enhancing efficiency in edge deployment.
- The method reduces parameter counts by factors of 1.80Γ to 3.82Γ and storage requirements by factors of 4.3Γ to 12.8Γ.
- Q-DEQ maintains competitive forecasting accuracy, with relative MSE differences ranging from -1.16% to +2.90% compared to explicit multi-layer baselines.
- The approach demonstrates the feasibility of using different combinatorial optimization backends for solving local QUBO problems.
Read more
Q-DEQ: Discrete Solving and Quantization for Deep Equilibrium Models in Time Series Forecasting under Edge Deployment Coding Constraints
Summary
This paper introduces Q-DEQ, a novel approach for enhancing deep equilibrium models (DEQs) in time series forecasting, particularly suited for edge deployment where parameter storage and computational efficiency are critical. Traditional DEQs utilize continuous optimization for updating coefficients, which can be inefficient for edge devices. Q-DEQ reformulates these updates as discrete optimization problems, leveraging local quadratic residual modeling and binary encoding to transform the coefficient search into a quadratic unconstrained binary optimization (QUBO) problem. This allows for the use of combinatorial optimization techniques such as simulated annealing (SA) and coherent Ising machines (CIM) for solving the updates. The proposed method integrates with a shared iTransformer backbone and employs W8A8 fake quantization to further reduce storage requirements. The evaluation on five multivariate time series datasets demonstrates that Q-DEQ achieves competitive forecasting accuracy compared to explicit multi-layer models while significantly reducing parameter counts and storage needs. The results indicate that local discrete solving can effectively enhance DEQ performance and provide a flexible framework for future applications in resource-constrained environments.
Methodology
Q-DEQ formulates local updates in DEQ forward solving as discrete optimization problems. It constructs candidate directions from the current state and iteration history, evaluates them using a local quadratic residual model, and encodes the direction coefficients into a QUBO problem. This problem can be solved using simulated annealing or a coherent Ising machine. A W8A8 fake quantization is applied in a re-forward pass to optimize weight and activation storage.
Results
The experiments show that Q-DEQ achieves relative MSE differences from the explicit multi-layer baseline ranging from -1.16% to +2.90%, with lower MSE on two datasets. The parameter sharing in DEQs leads to a reduction in parameter counts by factors of 1.80Γ to 3.82Γ, and combined with W8A8 quantization, static weight storage is reduced by factors of 4.3Γ to 12.8Γ. The results from CPU-based SA and CIM backend executions yield closely matching downstream forecasts.
Implications
The findings suggest that Q-DEQ can be effectively utilized in edge computing scenarios where resource constraints are prevalent. The integration of discrete optimization techniques into DEQ frameworks opens new avenues for efficient model deployment in real-time applications, particularly in time series forecasting.
Greedy Decoding Is Not Precision-Invariant: Cross-Precision Output Divergence in LLM Inference
Large Language Models
NLP
Theory
- Greedy decoding in LLMs is not precision-invariant, leading to different outputs across numerical formats.
- Up to 100% divergence was observed in outputs between BF16 and FP16 across various models and benchmarks.
- A predictive mechanism for output divergence was established, focusing on the top-two logit margin.
- Selective FP32 recomputation at critical points can improve output agreement significantly.
Read more
Greedy Decoding Is Not Precision-Invariant: Cross-Precision Output Divergence in LLM Inference
Summary
This paper investigates the phenomenon of cross-precision output divergence in large language models (LLMs) during greedy decoding, revealing that outputs can vary significantly between different numerical precisions (BF16 vs. FP16) even when using the same model and prompt. The authors conducted extensive evaluations across six models (ranging from 1.1B to 7B parameters) and three benchmarks, finding that 49-100% of prompts diverged, with a single token flip often leading to broader output discrepancies. They developed an empirical error-propagation analysis that identifies the conditions under which these divergences occur, particularly focusing on the top-two logit margin at the model's output layer. The paper proposes five testable predictions regarding interventions to mitigate these divergences, including a selective FP32 recomputation strategy that significantly improves output agreement with minimal latency overhead. The findings highlight the importance of precision stability during training and suggest that while the proposed interventions can reduce divergence, they do not guarantee deterministic outputs under all conditions.
Methodology
The authors conducted a systematic evaluation of output divergence across six LLMs using BF16 and FP16 formats. They employed an empirical error-propagation analysis to trace the source of divergences and tested five controlled interventions to validate their predictions about the conditions under which output discrepancies occur.
Results
The experiments demonstrated that a significant percentage of prompts diverged between BF16 and FP16, with the best-performing intervention (selective FP32 recomputation) achieving up to +36 percentage points in exact agreement with minimal latency overhead. The analysis confirmed that divergence is primarily influenced by the top-two logit margin relative to perturbations in the output layer.
Implications
The findings have important implications for the reproducibility and reliability of LLM outputs in practical applications, particularly in scenarios where different numerical formats may be used. The proposed interventions could enhance the consistency of model outputs across varying precision settings, improving the validity of evaluations and audits.
Muon Can Outperform Dedicated Continual Learning Methods
Optimization
Theory
Efficient ML
- Muon optimizer can effectively replace dedicated continual learning mechanisms in mitigating forgetting.
- One update-constraining mechanism suffices for effective continual learning; additional constraints do not enhance accuracy.
- Muon distributes updates across more effective singular directions compared to AdamW, enhancing learning efficiency.
- The study provides insights into the geometry of optimizer updates, suggesting it influences performance significantly.
Read more
Muon Can Outperform Dedicated Continual Learning Methods
Summary
This paper investigates the effectiveness of using the Muon optimizer in continual learning settings, specifically with Low-Rank Adapters (LoRA). The authors propose a plain incremental LoRA (IncLoRA) that utilizes Muon to orthogonalize updates, contrasting its performance against dedicated continual learning methods like O-LoRA and ELLA. The study reveals that IncLoRA combined with Muon achieves comparable or superior accuracy to these dedicated methods on the Standard CL Benchmark and shows improvements on TRACE. The findings suggest that a single update-constraining mechanism is sufficient, and the distribution of updates across singular directions is more critical than the magnitude of updates. The results challenge the notion that task-aware constraints are necessary for effective continual learning, indicating that the geometry of the optimizer's updates plays a significant role in mitigating catastrophic forgetting.
Methodology
The authors trained a plain incremental LoRA (IncLoRA) using the Muon optimizer, which orthogonalizes updates. They compared its performance against two dedicated continual learning methods, O-LoRA and ELLA, across multiple seeds and task orders on the Standard CL Benchmark and TRACE. The evaluation focused on overall accuracy, task fit, backward transfer, and forward transfer metrics.
Results
IncLoRA combined with Muon matched or exceeded the performance of dedicated methods like ELLA and O-LoRA. The analysis showed that Muon updates are distributed over 7.0 effective singular directions, while AdamW confines updates to 1.4 to 1.8 directions. The findings indicate that the geometry of updates, rather than their size, is crucial for performance in continual learning tasks.
Implications
The results suggest that optimizing update geometry can lead to more efficient continual learning models, potentially reducing the need for complex task-aware mechanisms. This could simplify the design of continual learning systems and enhance their applicability in real-world scenarios where model adaptability is critical.
Evaluating the Effectiveness of SechKAN on 1D Data
Theory
Efficient ML
Time Series
- SechKAN is a KAN variant using hyperbolic secant functions for 1D data.
- The model was evaluated on three benchmark datasets, showing competitive performance.
- Ablation studies highlighted the impact of grid size and normalization on model effectiveness.
- SechKAN's architecture allows for a reduced number of parameters comparable to MLPs.
Read more
Evaluating the Effectiveness of SechKAN on 1D Data
Summary
This paper investigates the effectiveness of SechKAN, a variant of Kolmogorov-Arnold Networks (KANs) that utilizes hyperbolic secant functions as basis functions, specifically for one-dimensional (1D) classification tasks. The study aims to fill the gap in the evaluation of SechKAN on 1D data, which has not been thoroughly explored despite its design for such inputs. The author evaluates SechKAN on three benchmark datasets: UCI Human Activity Recognition (UCI HAR), ElectricDevices, and Crop, comparing its performance against several established models, including EfficientKAN, MLP, CNN1D, ResNet1D, and DSCNN1D, all while maintaining comparable parameter budgets. The results demonstrate that SechKAN achieves competitive performance, particularly excelling on the Crop dataset. Additionally, ablation studies reveal that factors such as grid size and normalization significantly influence performance, indicating that the effectiveness of SechKAN is contingent on both the dataset and architectural choices. The paper contributes a systematic empirical evaluation of SechKAN, providing insights into its applicability and performance in the realm of 1D classification tasks.
Methodology
The methodology involves evaluating SechKAN on three 1D classification datasets and comparing its performance with other neural network architectures, including EfficientKAN, MLP, CNN1D, ResNet1D, and DSCNN1D. The evaluation includes ablation studies to assess the influence of grid size and normalization on performance.
Results
SechKAN demonstrated competitive performance across all three datasets, with particularly strong results on the Crop dataset. The ablation studies indicated that the choice of grid size and normalization techniques significantly affected the model's performance.
Implications
The findings suggest that SechKAN can be a viable alternative to traditional neural network architectures for 1D classification tasks, particularly in applications where parameter efficiency is crucial. The insights gained from this study may inform future research on KANs and their applications in various domains.
Trains but Doesn't Learn: A Post-Training Delivery Benchmark for LLM Agents as Forward-Deployed Engineers
Large Language Models
NLP
Theory
- Introduces the concept of TBDL, highlighting a significant failure mode in LLM agent delivery.
- Proposes a governed delivery plane with ten stages to evaluate agent performance beyond traditional metrics.
- Demonstrates that current LLM agents can avoid TBDL runs through adherence to delivery contracts.
- Establishes the importance of post-training delivery evaluation in the context of LLM deployment.
Read more
Trains but Doesn't Learn: A Post-Training Delivery Benchmark for LLM Agents as Forward-Deployed Engineers
Summary
This paper introduces a novel evaluation framework for large language model (LLM) agents functioning as forward-deployed engineers (FDEs) in post-training delivery scenarios. The authors highlight a critical failure mode termed 'trains but does not learn' (TBDL), where an agent optimizes metrics without delivering a model that meets customer needs. The study proposes a governed delivery plane consisting of ten stages, each evaluated by an oracle based on platform-recorded facts, to ensure accountability in the delivery process. The authors conducted experiments with four leading LLM agents across various GPU configurations, demonstrating that current frontier agents successfully avoided TBDL runs by adhering to a structured delivery contract. The findings emphasize the importance of evaluating agents on their delivery capabilities rather than mere performance metrics, showcasing the potential for improved trust in LLM agents for real-world applications.
Methodology
The authors developed a governed delivery plane consisting of ten stages, where each stage is scored by an oracle based on factual data from the platform. They conducted end-to-end evaluations of four frontier LLM agents across various GPU configurations, ensuring that each scenario passed both positive and negative controls before scoring.
Results
In their experiments, the authors found that all 38 certified real episodes produced zero TBDL deliveries and zero gate bypasses. The agents successfully adhered to the delivery contract, demonstrating that they could avoid the critical failure mode of delivering models that did not meet customer requirements. The study also revealed that the agents performed well under pressure, maintaining governance in all trials.
Implications
The findings suggest that LLM agents can be trusted to deliver effective models in real-world applications, provided they operate within a structured delivery framework. This has implications for industries relying on LLMs for critical tasks, such as finance and healthcare, where the accuracy and reliability of delivered models are paramount.
AirGC-CD: Gaussian-Circulant Precoding for Exactly Debiasable PAPR Reduction in Over-the-Air Federated Learning
Federated Learning
Optimization
Efficient ML
- Introduces AirGC-CD for reducing PAPR in over-the-air federated learning.
- Utilizes Gaussian-circulant precoding to ensure unbiased aggregation.
- Achieves significant compression of model updates, reducing transmission load.
- Derives a near-optimal clipping ratio balancing variance and SNR loss.
Read more
AirGC-CD: Gaussian-Circulant Precoding for Exactly Debiasable PAPR Reduction in Over-the-Air Federated Learning
Summary
The paper introduces AirGC-CD, a novel scheme for over-the-air federated learning (FL) that addresses the challenges of high peak-to-average power ratio (PAPR) in local model updates transmitted by edge devices. Traditional methods for reducing PAPR, such as clipping and power back-off, introduce biases or degrade signal quality. AirGC-CD employs a Gaussian-circulant precoding technique that ensures the output distribution is Gaussian, allowing for effective clipping without introducing bias. This method reduces the transmission load from the model dimension to a sketch dimension, significantly compressing the data while maintaining unbiased aggregation. The authors derive a near-optimal clipping ratio that balances variance and signal-to-noise ratio (SNR) loss. The convergence analysis under peak power constraints shows an O(1/βT) convergence bound without a bias floor. Experimental results demonstrate that AirGC-CD outperforms existing over-the-air FL methods, particularly in low SNR scenarios, while using fewer channel resources per round.
Methodology
The authors propose a Gaussian-circulant precoding method that transforms local model updates into a Gaussian distribution before clipping. This transformation allows for the clipping distortion to be inverted on average, resulting in unbiased aggregation. The method also compresses the transmission from model dimension d to sketch dimension m using fast Fourier transforms, optimizing both communication efficiency and signal integrity.
Results
AirGC-CD was shown to outperform baseline over-the-air federated learning schemes in various settings, especially under low SNR conditions. The method effectively reduces the number of channel uses required per round while maintaining convergence rates and ensuring unbiased aggregation.
Implications
The findings suggest that AirGC-CD can facilitate more efficient federated learning in environments with limited communication resources, such as mobile networks and IoT applications. This could lead to broader adoption of federated learning in real-world applications where data privacy and communication efficiency are critical.
CoEvo: Oracle-Grounded Self-Evolution of a Single Model for Multi-Step Causal Reasoning
NLP
Large Language Models
Theory
- CoEvo framework enables self-evolution of LLMs for multi-step causal reasoning.
- Utilizes an oracle to verify individual reasoning steps, enhancing model supervision.
- Model alternates between Proposer and Solver roles to generate and validate reasoning chains.
- Achieves superior performance on multi-step reasoning benchmarks compared to traditional methods.
Read more
CoEvo: Oracle-Grounded Self-Evolution of a Single Model for Multi-Step Causal Reasoning
Summary
The paper presents CoEvo, an innovative framework for enhancing multi-step causal reasoning in large language models (LLMs) by leveraging an oracle for self-evolution. Multi-step causal reasoning involves chaining inferences where each step is dependent on the previous one, making it susceptible to error propagation. Traditional methods like knowledge distillation are inadequate due to the inherent flaws in teacher models and the need for privacy in specialized domains. CoEvo addresses these challenges by allowing a single model to alternate between two roles: Proposer and Solver. As a Solver, the model generates competing reasoning chains, while an oracle verifies the correctness of individual steps, providing a form of supervision. As a Proposer, the model creates increasingly complex scenarios within the oracle's constraints, facilitating a curriculum that promotes deeper reasoning. The framework enables the model to adaptively improve its reasoning capabilities while maintaining accuracy in unseen categories. The results demonstrate that CoEvo significantly outperforms traditional distillation methods and achieves high path correctness, showcasing its effectiveness in real-world applications.
Methodology
The CoEvo framework employs a dual-role approach where the model acts as both Proposer and Solver. The Proposer generates complex reasoning scenarios within the constraints of an oracle, while the Solver produces competing reasoning chains that the oracle verifies. This process allows for adaptive task generation and co-evolution of training pressure alongside model capabilities.
Results
The CoEvo framework was tested on industrial, clinical, and legal multi-step causal reasoning benchmarks, where an 8B LLM achieved a path correctness rate of 82.1%, significantly surpassing the 71.4% achieved by the strongest proprietary reference and traditional distillation baselines. The model also demonstrated robust generalization to unseen categories and systems.
Implications
The CoEvo framework has significant implications for deploying LLMs in high-stakes domains such as industrial fault diagnosis, clinical decision-making, and legal analysis, where reliable multi-step causal reasoning is critical. By enabling self-evolution and adaptive learning, it enhances the practicality and effectiveness of LLMs in real-world applications.
CSC: Calibrated Simplicity for Conflict-Aware Social Bot Detection in the LLM Era
Graph Learning
NLP
Multimodal
- Introduces CSC, a framework for detecting social bots that addresses modality conflict in the LLM era.
- Combines simplified graph-based detection with calibrated confidence alignment and inconsistency modeling.
- Demonstrates improved decision quality and reliability in bot detection through experimental validation.
- Highlights the importance of structural signals in bot detection amidst semantic camouflage.
Read more
CSC: Calibrated Simplicity for Conflict-Aware Social Bot Detection in the LLM Era
Summary
The paper addresses the challenge of social bot detection in the era of large language models (LLMs), which have made it easier for bots to camouflage their identities through fluent and human-like text. This creates a modality conflict where textual evidence may appear benign while structural evidence from graphs and profiles suggests otherwise. The authors propose a novel framework called Calibrated Simplicity for Conflict-Aware social bot detection (CSC), which emphasizes a balance between simplicity and effectiveness. The CSC framework integrates three main components: a simplified prototype-guided graph expert that retains useful structural biases, calibrated simplex-constrained fusion for aligning heterogeneous confidence scores, and a lightweight inconsistency expert that models disagreements across modalities. The experiments conducted on datasets such as TwiBot-22 and MGStBot-large demonstrate that CSC enhances decision quality while maintaining competitiveness across benchmarks. The findings indicate that calibration improves confidence reliability, and the inconsistency expert provides localized corrections in high-conflict scenarios. Overall, the results suggest that preserving reliable structural signals and calibrating conflicting modalities is more beneficial than merely increasing complexity in graph-based detection systems.
Methodology
The CSC framework consists of three components: a simplified prototype-guided graph expert that leverages structural biases, calibrated simplex fusion that aligns confidence scores from different modalities, and an inconsistency expert that corrects for disagreements among the modalities. The methodology emphasizes maintaining simplicity while effectively addressing the challenges posed by LLMs in social bot detection.
Results
Experiments on multiple datasets (TwiBot-22, TwiBot-20, MGStBot-large) show that CSC improves the quality of calibrated operating-point decisions. The framework outperforms existing methods in terms of reliability and stability, particularly in high-conflict scenarios where textual evidence is camouflaged. The analysis reveals that the inconsistency expert effectively corrects localized discrepancies, and the simplified graph-side control enhances the stability-cost trade-off.
Implications
The findings suggest that social media platforms can enhance their bot detection capabilities by adopting frameworks like CSC that prioritize structural integrity and confidence calibration. This approach could lead to more robust defenses against misinformation and manipulation in online discourse, particularly as bot technology continues to evolve.
Not All Ranks Are Equal: Budget-Aware LoRA Merging Across Tasks
Efficient ML
Computer Vision
NLP
- Identifies uniform rank allocation as a major source of performance gaps in LoRA merging.
- Introduces Net Utility, a data-free metric for optimal rank allocation based on task utility and interference.
- Demonstrates improved performance across multiple merging methods and tasks using the proposed allocation strategy.
- Achieves an average performance improvement of +2.1% in vision tasks and +2.2% in language tasks.
Read more
Not All Ranks Are Equal: Budget-Aware LoRA Merging Across Tasks
Summary
This paper addresses the inefficiencies in merging low-rank adapters (LoRAs) for multi-task learning, particularly the assumption that all layers and tasks should share the same rank budget. The authors argue that this uniform rank allocation leads to significant performance gaps between merged and per-task LoRAs. To tackle this issue, they introduce a novel metric called Net Utility, which evaluates the utility of singular directions derived from the Singular Value Decomposition (SVD) of each task's LoRA. This metric scores each direction based on its contribution to its own task and its interference with other tasks, allowing for a more informed and effective allocation of rank budgets. The proposed method is tested across various merging methods and tasks in both vision and language domains, demonstrating that the Net Utility approach significantly enhances performance compared to traditional uniform allocation methods.
Methodology
The authors propose a greedy approach to rank allocation using the Net Utility metric, which scores singular directions from the SVD of task-specific LoRAs. This scoring considers both the utility for the specific task and the potential interference with other tasks. The top-scoring directions are selected while adhering to a global budget constraint, allowing for a more tailored merging process.
Results
The Net Utility-based rank allocation outperforms traditional uniform allocation methods, achieving an average performance improvement of +2.1% on vision tasks and +2.2% on language tasks. Some method-space combinations even show improvements as high as +3.8%.
Implications
The findings suggest that more nuanced rank allocation strategies can lead to better performance in multi-task learning scenarios, particularly in resource-constrained environments like on-device applications. This could enhance the deployment of models that require efficient parameter usage without sacrificing accuracy.
Iterative Atom Refinement: A Monotonicity Principle for Dictionary Learning
Theory
Optimization
Efficient ML
- Introduction of the Iterative Atom Refinement (IAR) algorithm for dictionary learning.
- Development of a rigorous convergence theory based on a new monotonicity principle.
- Demonstration that IAR can identify a generating atom after only three refinement steps.
- Numerical experiments validate the theoretical results and show rapid convergence.
Read more
Iterative Atom Refinement: A Monotonicity Principle for Dictionary Learning
Summary
This paper introduces the Iterative Atom Refinement (IAR) algorithm, a novel approach for dictionary learning that aims to recover individual dictionary atoms from sparse observations. The authors present a rigorous convergence theory for IAR, which operates by iteratively selecting observations that are most correlated with the current estimate of the dictionary atom and refining this estimate through averaging. A key contribution of the paper is the identification of a monotonicity principle for atom-selection probabilities, which ensures that a small initial advantage in selecting the correct atom is amplified through the refinement process. This principle allows IAR to successfully identify a generating atom after just three iterations, contrasting with traditional methods that require closer initialization to the true dictionary. The authors provide theoretical guarantees for the algorithm's performance under high-dimensional probabilistic models and support their findings with numerical experiments that demonstrate rapid convergence even with partial data and nonorthogonal dictionaries. Overall, IAR represents a significant advancement in the field of dictionary learning, particularly in its ability to recover dictionaries from random initialization without the need for complex optimization techniques.
Methodology
The IAR algorithm begins with a random initialization and iteratively refines its estimate of dictionary atoms by selecting observations that exhibit the strongest correlations with the current estimate. The selected observations are averaged and projected back onto the unit sphere to generate the next iterate. The analysis introduces atom-selection frequencies to track the proportion of selected samples containing each dictionary atom, leveraging a monotonicity principle to ensure that the ordering of these frequencies is preserved throughout the refinement process.
Results
Theoretical results show that IAR can identify the dictionary atom with the highest initial correlation after three iterations with high probability. Additionally, the algorithm demonstrates effective recovery of the entire dictionary from random initializations, provided that the number of independent runs is sufficiently large. Numerical experiments confirm the rapid convergence of IAR for both partial data and nonorthogonal dictionaries.
Implications
The findings suggest that IAR could be a powerful tool for applications in image processing and other fields where dictionary learning is essential. Its ability to recover dictionaries from random initialization may simplify the implementation of dictionary learning in practical scenarios, reducing reliance on careful initialization strategies.
Continuous Optimization for p-adic Models
Optimization
Theory
- Introduces continuous optimization methods for p-adic models, overcoming limitations of discrete searches.
- Utilizes the Berkovich affine line to create a path-connected space for p-adic numbers.
- Demonstrates effective gradient descent algorithms, including Momentum and Adam, for p-adic parameters.
- Shows that p-adic models can perform tasks like modular arithmetic and binary-encoded classification.
Read more
Continuous Optimization for p-adic Models
Summary
This paper introduces a novel method for continuous gradient descent applicable to machine learning models with p-adic parameters, addressing a significant gap in existing optimization techniques that primarily rely on discrete or combinatorial searches. The authors leverage the Berkovich affine line, a path-connected expansion of the p-adic numbers Qp, to enable continuous optimization. This approach allows for the formulation of effective gradient descent algorithms, including Momentum and Adam variants, specifically tailored for p-adic models. The paper demonstrates that these continuous optimizers can efficiently learn linear models with p-adic coefficients, which are capable of performing modular arithmetic tasks that traditional real-valued models cannot handle. Additionally, the authors provide a comprehensive framework for backpropagation in p-adic networks and release an open-source library to facilitate further research in this area. The work resolves several open problems in the field and opens new avenues for developing deep neural networks with p-adic parameters.
Methodology
The authors construct a canonical path-connected space for p-adic numbers using the Berkovich affine line, enabling the definition of continuous losses and informative directional derivatives. They develop continuous optimizers through grouped coordinate descent and formulate backpropagation for training p-adic models.
Results
The proposed methods successfully train linear models with p-adic coefficients for tasks such as modular arithmetic and classification on binary-encoded hierarchies. The continuous optimization framework demonstrates improved performance over traditional discrete methods, effectively addressing open problems in the field.
Implications
This research has the potential to enhance machine learning applications that involve data with p-adic or ultrametric structures, such as certain types of hierarchies and number-theoretic algorithms. It paves the way for the development of deep neural networks that can leverage the unique properties of p-adic numbers.
Towards Hierarchical GNNs for multi-grid power flow: generalization across operating scenarios
Graph Learning
- Introduction of hierarchical latent communication to improve multi-grid power-flow modeling.
- Kron reduction significantly outperforms flat models in reducing voltage error.
- Demonstrated generalization to new operating scenarios within training topologies.
- Models do not yet achieve cross-topology generalization, indicating a future research direction.
Read more
Towards Hierarchical GNNs for multi-grid power flow: generalization across operating scenarios
Summary
This paper presents a novel approach to improving the generalization of multi-grid power-flow models using hierarchical graph neural networks (GNNs). The authors introduce a hierarchical latent communication module that enhances the exchange of information through two reduced graphs within a GENCO-based corrective network. The study compares different transport constructions, including Kron-derived weights and a graph quotient, across three grid topologies. The results indicate that the Kron reduction significantly reduces the macro family-balanced voltage error, achieving an 85% reduction compared to a flat backbone model and a 31% reduction compared to the Quotient model. Both hierarchical models consistently outperform a per-bus mean fitted on training solutions across all training topologies. However, while the models demonstrate effective generalization to new operating scenarios, they do not yet outperform fitted references in cross-topology evaluations. This work lays the groundwork for future developments aimed at achieving generalization across unseen topologies.
Methodology
The authors implemented a hierarchical communication module within the GENCO solver, utilizing two transport constructions: Kron-derived weights and a graph quotient. They conducted preliminary training over 200 epochs on three grid topologies, evaluating model performance on 200 newly generated scenarios for each grid. The evaluation distinguished between scenario generalization (fixed topology) and topology generalization (new grid absent from optimization).
Results
The Kron reduction model achieved a macro family-balanced voltage error reduction from 5.660 Β± 0.899 to 0.851 Β± 0.110, marking an 85% improvement over the flat backbone and a 31% improvement over the Quotient model. Both hierarchical models outperformed the per-bus mean across all training topologies and initialization seeds. However, the models did not surpass fitted references in cross-topology evaluations, highlighting the need for further development.
Implications
The findings suggest that hierarchical GNNs can effectively enhance the modeling of multi-grid power flows, potentially leading to improved performance in real-world applications such as power system management and optimization. Future work could focus on achieving generalization across different grid topologies, which would further enhance the applicability of these models.
CLOOPD: Closing the Learner Loop in On-Policy Distillation
Reinforcement Learning
Optimization
Efficient ML
- CLOOPD separates teacher-signal acquisition from student-side realization, optimizing the learning process.
- The framework allows for multiple passes over the same teacher-scored batch, improving efficiency and effectiveness.
- CLOOPD policies achieve higher macro accuracy compared to traditional methods while using fewer resources.
- Adaptive Ξ± waypoint selection is crucial for maintaining safe and effective learning boundaries.
Read more
CLOOPD: Closing the Learner Loop in On-Policy Distillation
Summary
The paper introduces CLOOPD, a novel framework for On-Policy Distillation (OPD) that enhances the efficiency of learning by separating the acquisition of teacher signals from the realization process on the student side. Traditional OPD methods incur high costs by requiring the student to generate new trajectories and the teacher to score them in a tightly coupled manner. CLOOPD addresses this by freezing the teacher-scored batch and its advantages, allowing for multiple passes of the student over the same data without additional teacher calls. The framework employs adaptive waypoint selection within a KL envelope and includes various policies such as CLOOPD-Fixed2 and CLOOPD-Fixed3, which optimize the use of teacher tokens and actor resources. Experimental results demonstrate that CLOOPD policies significantly outperform the one-pass TOP-D anchor in terms of macro accuracy while using fewer teacher tokens and GPU hours. The findings suggest that CLOOPD effectively closes the learner loop, enabling deeper realization of teacher signals and better resource management in OPD.
Methodology
CLOOPD employs a closed-loop framework consisting of three stages: Probe, Observe, and Allocate. It constructs adaptive waypoints for teacher signals, freezes the scored batch for multiple student passes, and allocates actor tokens based on realization policies. The framework includes various configurations such as CLOOPD-Fixed2, CLOOPD-Fixed3, and CLOOPD-TPMR, which are tested across multiple experimental runs to evaluate their performance.
Results
In experiments involving six 300-step runs, CLOOPD-Fixed3 achieved a macro accuracy of 19.36, significantly higher than the 15.41 achieved by the one-pass TOP-D method. CLOOPD-Fixed3 also matched TOP-D's performance at step 300 by step 100 while using 67.2% fewer teacher-scored tokens and 28.0% fewer GPU hours. The results indicate a strong realization-depth frontier and effective teacher-token amortization.
Implications
CLOOPD's approach to optimizing resource allocation in OPD has potential applications in various reinforcement learning scenarios, particularly where efficient use of teacher signals is critical. The framework can inform future research on improving the scalability and effectiveness of distillation methods in machine learning.
From Experts to Sub-experts: Fine-grained Parameter-Efficient Fine-Tuning for MoE LLMs
Large Language Models
Efficient ML
Optimization
- NSFT offers a fine-grained approach to parameter-efficient fine-tuning, moving beyond expert-level adaptations.
- The method utilizes structured channel groups within experts to select task-relevant sub-experts for tuning.
- NSFT incorporates adaptive training strategies to compensate for reduced update magnitudes during sparse tuning.
- Experimental results show NSFT achieves superior performance with fewer trainable parameters compared to existing methods.
Read more
From Experts to Sub-experts: Fine-grained Parameter-Efficient Fine-Tuning for MoE LLMs
Summary
This paper addresses the challenges of parameter-efficient fine-tuning (PEFT) in large language models (LLMs), particularly those using Mixture-of-Experts (MoE) architectures. As LLMs grow in size, traditional dense full-parameter adaptation becomes computationally expensive, prompting the need for more efficient tuning methods. The authors propose a novel framework called Neural Sub-expert Fine-Tuning (NSFT), which refines the adaptation process from the expert level to a finer granularity of sub-experts. NSFT identifies task-relevant sub-experts by analyzing channel activations within experts, allowing for more precise updates that avoid the redundancy of updating entire experts. The method incorporates learning-rate scaling and dynamic gradient scaling to optimize the sparse updates. Experimental results demonstrate that NSFT outperforms existing PEFT methods, achieving better accuracy with significantly fewer trainable parameters across various domain-specific tasks and benchmarks. This suggests that sub-expert-level adaptation is a more effective and efficient approach for fine-tuning MoE LLMs.
Methodology
The NSFT framework decomposes each expert into structured channel groups and selects relevant sub-experts based on routing importance and activation saliency. It employs learning-rate scaling and dynamic gradient scaling to enhance the effectiveness of sparse updates.
Results
NSFT consistently outperformed existing PEFT and expert-level sparse tuning methods across multiple tasks, achieving up to 66.9% reduction in trainable parameters while maintaining or improving accuracy. The method demonstrated superior accuracy-parameter trade-offs compared to LoRA and ESFT.
Implications
The proposed NSFT method could lead to more efficient training of large language models, enabling their deployment in resource-constrained environments. It also opens avenues for further research into adaptive tuning strategies and the exploration of sparsity in model architectures.
Fast Matrix Multiplication in fp8: Certified Coefficient Optimization and Measured Error
Optimization
Efficient ML
Theory
- Introduces a coefficient functional Ξ¦ to optimize matrix multiplication in fp8.
- Certifies the global optimum of Ξ¦ using geometric methods on a Hadamard manifold.
- Demonstrates that the optimal realization reduces fp8 error significantly compared to classic Strassen algorithms.
- Establishes a connection between coefficient geometry and expected error in fp8.
Read more
Fast Matrix Multiplication in fp8: Certified Coefficient Optimization and Measured Error
Summary
This paper addresses the optimization of fast matrix multiplication algorithms in the fp8 numerical format, focusing on the Strassen-type algorithms. The authors introduce a coefficient functional, Ξ¦, which summarizes the coefficient geometry of different algorithm realizations. They minimize this functional over the change-of-basis orbit to certify the global optimum, establishing a connection between the coefficient geometry and the expected error in fp8. The study reveals that every exact real rank-7 realization has a predicted RMS constant that is at least 5/3 times that of the cubic algorithm, given a fixed noise coefficient. The authors also present a block-scaled e4m3 model where Ξ¦ serves as the leading-order coefficient of the expected mean-squared error. Experimental results demonstrate that the Ξ¦-predicted ordering aligns with the actual fp8 error, confirming that the optimal realization significantly reduces the excess negative log-likelihood (NLL) in large models. This work transforms the choice of algorithm realization into a mathematically certified design problem, enhancing the reliability of low-precision matrix multiplication in practical applications.
Methodology
The authors employ a geometric approach to minimize the coefficient functional Ξ¦ over the change-of-basis orbit, treating it as a KempfβNess problem on a Hadamard manifold. They derive a second-moment model for fused block-scaled fp8 to establish the relationship between Ξ¦ and expected mean-squared error, followed by empirical testing of the predicted ordering against actual fp8 errors.
Results
The study finds that the Ξ¦-optimal realization consistently lowers measured fp8 error across various tested decompositions, achieving reductions of at least 1.14 times. In practical applications involving large models, the optimal realization reduces the excess NLL by 10 to 55% compared to traditional Strassen methods.
Implications
This research has significant implications for the design of efficient matrix multiplication algorithms in low-precision formats, particularly in large-scale machine learning applications. The certified optimization approach can enhance the accuracy and reliability of computations in environments where fp8 is commonly used.
Task-Aware Hybrid QUBO Optimization for Structured Neural Network Pruning
Optimization
Efficient ML
Computer Vision
- Introduces a hybrid global-local framework for mixed-precision quantization.
- Utilizes Task-Aware QUBO to model quantization error and layer sensitivity.
- Incorporates graph-aware constraints to manage activation precision compatibility.
- Employs PROTES for local refinement of precision assignments.
Read more
Task-Aware Hybrid QUBO Optimization for Structured Neural Network Pruning
Summary
This paper presents a novel hybrid global-local framework for task-aware mixed-precision quantization of deep neural networks, addressing the combinatorial optimization problem of layer-wise bit-width assignment. The authors formulate a Task-Aware QUBO (Quadratic Unconstrained Binary Optimization) that incorporates measured quantization error and layer sensitivity to optimize precision assignments. They introduce graph-aware structural constraints to ensure compatible activation precisions at critical junctions and to mitigate high-damage assignments in adjacent paths. The framework is initialized with a high-quality QUBO solution and refined using PROTES, a probabilistic optimization method that evaluates the network's performance directly. The experiments conducted on a compact image-denoising model, HalfUNet, demonstrate significant improvements in accuracy-compute and accuracy-storage trade-offs compared to uniform and unconstrained routing baselines, highlighting the effectiveness of the proposed method in optimizing neural network performance under resource constraints.
Methodology
The authors develop a Task-Aware QUBO framework that models quantization damage and layer sensitivity. They implement graph-aware constraints to enhance the optimization process and utilize the PROTES method for refining the precision assignments based on direct network evaluation.
Results
The proposed method shows improved accuracy-compute and accuracy-storage trade-offs when tested on the HalfUNet model against existing uniform and unconstrained routing methods, indicating its effectiveness in optimizing neural network configurations.
Implications
This research has significant implications for deploying deep learning models on resource-constrained edge devices, as it enables efficient mixed-precision quantization that balances performance and computational cost, potentially enhancing the feasibility of high-fidelity models in practical applications.
Gaussian Flow-Matching Schedules: Implications for Sampling and Training
Generative Models
Theory
Efficient ML
- Introduces a path-factorization decomposition for flow-matching schedules.
- Establishes necessary conditions for exact finite-step Euler sampling.
- Derives closed-form solutions to minimize or equalize regression variance.
- Demonstrates the separation of sampling dynamics and regression variance control.
Read more
Gaussian Flow-Matching Schedules: Implications for Sampling and Training
Summary
This paper investigates the impact of flow-matching schedules on both sampling dynamics and regression variance in the context of Gaussian distributions. The authors introduce a path-factorization decomposition that separates the design choices into a variance path, which dictates the intermediate distributions and probability flow, and a mixing angle that controls the regression variance without altering the flow. They derive necessary conditions for exact finite-step Euler sampling and provide closed-form solutions that either minimize or stabilize the regression variance along a specified path. The findings highlight the distinct roles of path design in influencing both the efficiency of sampling and the complexity of training, suggesting that optimizing these components can lead to improved performance in flow-matching applications.
Methodology
The authors analyze Gaussian flow-matching by defining a stochastic interpolant and exploring the effects of different schedules on both sampling and training. They utilize mathematical formulations to derive the variance path and mixing angle, and they assess the implications of these components on the accuracy of finite-step Euler sampling and regression variance.
Results
The study reveals that the flow-matching drift can be decomposed into components that independently influence sampling and regression variance. The authors provide explicit conditions for achieving exact sampling and demonstrate how different design choices can lead to either minimized or constant regression variance along the flow path.
Implications
The findings suggest that careful design of flow-matching schedules can enhance the efficiency of sampling processes and reduce the complexity of training in machine learning models, particularly those involving Gaussian distributions. This has potential applications in generative modeling and other areas where flow-based methods are employed.
PROSE: A Theory of Optimal Stopping with Perishable Evidence for Peer Selection in Intermittently Connected Decentralised Learning
Federated Learning
Theory
Optimization
- Introduces a theory of optimal stopping for peer selection in decentralized federated learning.
- Develops a reservation-value threshold rule for optimal decision-making under perishable evidence.
- Establishes confidence-safe certification and mobility-aware value-of-information rules.
- Demonstrates that a one-step confidence-safe rule is optimal in volatile mobility regimes.
Read more
PROSE: A Theory of Optimal Stopping with Perishable Evidence for Peer Selection in Intermittently Connected Decentralised Learning
Summary
This paper presents a theoretical framework for optimal stopping in the context of peer selection during decentralized federated learning, particularly in intermittently connected mobile systems. The author identifies the unique challenges posed by perishable evidence, where the quality of peer information decays due to transient connectivity and model drift. The study formulates the peer-selection decision as a finite-horizon Markov optimal-stopping problem, introducing a reservation-value threshold rule for optimal policy determination. Key contributions include the development of confidence-safe certification methods, mobility-aware value-of-information rules, and a closed-form expression for the value of waiting under marked-Poisson contact arrivals. The paper also establishes a myopic-optimality theorem, demonstrating that in volatile mobility conditions, a one-step confidence-safe rule serves as an effective surrogate for the optimal policy. The framework is instantiated as PROSE (Perishable-evidence Reservation-value Optimal Stopping for Exchange), a lightweight, local policy that recovers classical decision-making limits under specific conditions. Overall, the work provides a comprehensive analytical approach to understanding when and how model exchanges should occur in decentralized learning environments with perishable evidence.
Methodology
The methodology involves formulating the peer-selection problem as a finite-horizon Markov optimal-stopping problem, analyzing the decision-making process through dynamic programming techniques. The author derives various rules and bounds analytically, focusing on the implications of mobility and perishable evidence on peer selection.
Results
The paper proves the existence of an optimal policy characterized by a reservation value, establishes confidence bounds for peer selection, and provides a closed-form value of waiting under specific contact arrival conditions. It also shows that higher link hazards reduce the value of continued probing, thus enlarging the stopping region.
Implications
The findings have significant implications for decentralized federated learning systems, particularly in mobile and dynamic environments. The proposed framework can enhance peer selection strategies, improve model exchange efficiency, and inform the design of decentralized learning protocols that account for transient connectivity and model drift.
Correcting Within-Group Self-Selection Bias in Prioritized Replay
Reinforcement Learning
- Identifies and quantifies the issue of within-group self-selection bias in Prioritized Experience Replay.
- Proposes sibling-aware replay methods that separate group-level priority mass from within-group outcome selection.
- Demonstrates improved learning efficiency in exact state-action environments and in function-approximation settings.
- Introduces three methods: SAMPLE, AVG, and MODEL, which utilize empirical sibling distributions for better learning outcomes.
Read more
Correcting Within-Group Self-Selection Bias in Prioritized Replay
Summary
This paper addresses the issue of within-group self-selection bias in Prioritized Experience Replay (PER), a technique that enhances sample efficiency in reinforcement learning by replaying high-priority transitions. The authors identify that in stochastic environments, PER can distort the distribution of outcomes from transitions with the same state-action pair, leading to biased learning. They propose a decomposition of PER into between-group allocation and conditional sibling selection, allowing for fixed-buffer corrections that maintain the priority mass at the group level. The proposed methods, SAMPLE, AVG, and MODEL, utilize sibling-aware replay to improve learning efficiency by focusing on empirical outcome frequencies within selected state-action groups. The authors demonstrate that these methods outperform traditional PER in exact state-action environments and mitigate performance degradation in various games within the MinAtar framework. Overall, the paper formalizes a new approach to experience replay that retains the benefits of prioritization while correcting for biases in outcome selection.
Methodology
The authors decompose Prioritized Experience Replay into two components: between-group allocation and conditional sibling selection. They propose three methodsβSAMPLE, AVG, and MODELβthat utilize sibling-aware replay to correct for biases in outcome selection while preserving the group-level priority mass. These methods are evaluated in both exact state-action environments and function-approximation settings using frozen VQ-VAE codes.
Results
The proposed sibling-aware replay methods showed improved learning efficiency over traditional PER in exact state-action environments. In the MinAtar framework, the SAMPLE method effectively mitigated performance degradation under mean-preserving reward tails in four out of five games, demonstrating the effectiveness of the sibling-aware approach.
Implications
The findings suggest that correcting for within-group self-selection bias can lead to more efficient learning in reinforcement learning applications, particularly in stochastic environments. This approach could be beneficial for developing more robust reinforcement learning agents that can better handle variability in outcomes.
CacheDyG: Decoupling Temporal Propagation for Efficient Dynamic Graph Learning
Graph Learning
Efficient ML
Optimization
- CacheDyG decouples temporal propagation from parameter updates, reducing computational redundancy.
- The framework utilizes a Temporal Dependency Cache to store reusable node-time representations.
- CacheDyG achieves competitive predictive performance with significantly fewer trainable parameters.
- The method demonstrates lower runtime and improved scalability on larger datasets.
Read more
CacheDyG: Decoupling Temporal Propagation for Efficient Dynamic Graph Learning
Summary
The paper introduces CacheDyG, a novel framework designed to enhance the efficiency of dynamic graph learning by decoupling temporal propagation from routine parameter updates. Traditional dynamic graph neural networks often recompute historical structures during every optimization epoch, leading to significant computational overhead and increased memory usage due to large trainable representations for each node-time pair. CacheDyG addresses these inefficiencies by implementing a Temporal Dependency Cache that stores non-trainable, graph-aware node-time representations. During training, the model utilizes these cached representations, updating only a lightweight cache refiner and a link predictor. This approach minimizes redundant computations and reduces the number of trainable parameters while maintaining competitive predictive performance. The authors validate CacheDyG across five dynamic graph benchmarks, demonstrating its superior efficiency and effectiveness compared to existing methods.
Methodology
CacheDyG employs a cache-refine framework that constructs a Temporal Dependency Cache to store historical graph representations. It performs lightweight updates to a cache refiner and a link predictor during training, while selectively refreshing the cache to align with supervised objectives. This decoupling allows for amortized representation construction, focusing on optimizing cached signals rather than recomputing historical structures.
Results
CacheDyG outperforms baseline methods in terms of average precision on all evaluated datasets, achieving the best performance with only 19.106K trainable parameters. It also demonstrates the lowest runtime, making it feasible for larger datasets where other memory-intensive models struggle. The results highlight a favorable accuracy-efficiency trade-off, supported by ablation and sensitivity analyses.
Implications
The findings suggest that CacheDyG can be applied to various real-world applications involving dynamic graphs, such as social networks, recommender systems, and biological interaction networks. Its efficient learning framework can facilitate the analysis of large-scale, time-evolving relational systems without incurring prohibitive computational costs.
Multi-View Fair Clustering Guided by Cross-View Sensitive Information Discrepancy
Theory
Optimization
Multimodal
- Introduces a novel framework for multi-view fair clustering that considers cross-view sensitive information discrepancies.
- Employs a bias-ranked asymmetric alignment mechanism to regulate learning between views with differing levels of sensitive dependence.
- Imposes fairness regularization on clustering assignments to promote group fairness.
- Achieves competitive performance compared to state-of-the-art methods while maintaining stability under varying sensitive information levels.
Read more
Multi-View Fair Clustering Guided by Cross-View Sensitive Information Discrepancy
Summary
This paper addresses the challenge of fair clustering in multi-view settings, where data is represented from multiple perspectives. While existing methods have improved clustering performance, they often overlook the varying levels of sensitive information encoded in different views. This oversight can lead to biased clustering outcomes, particularly in sensitive applications. The authors propose a novel framework that explicitly models cross-view sensitive information discrepancies. By estimating the sensitive dependence of each view, they introduce a bias-ranked asymmetric alignment mechanism that encourages views with higher sensitive dependence to learn from those with lower dependence. This approach is complemented by fairness regularization on consensus soft assignments to enhance group fairness. The proposed method is evaluated on benchmark datasets, demonstrating a favorable balance between clustering quality and fairness, as well as robustness to heterogeneous sensitive dependence across views.
Methodology
The methodology involves using view-specific autoencoders to learn latent representations for each view, quantifying the sensitive dependence of each view, and implementing an asymmetric alignment mechanism that regulates cross-view learning based on these dependencies. Fairness regularization is applied to the consensus soft assignments to ensure equitable treatment across demographic groups.
Results
The proposed framework shows significant improvements in balancing clustering quality and group fairness across five benchmark multi-view datasets. It outperforms existing multi-view clustering and fairness-aware methods, particularly in scenarios with heterogeneous sensitive information, maintaining stable fairness while achieving competitive clustering performance.
Implications
The findings suggest that incorporating sensitivity discrepancies into multi-view clustering can lead to more equitable outcomes in socially sensitive applications, such as healthcare and finance, where clustering decisions can impact demographic groups differently.
Hill Sampling for Test-Time Scaling: A Simple and Better Alternative to Repeated Sampling, Evolution, and Training
Large Language Models
Optimization
- Hill Sampling achieves state-of-the-art performance on circle packing and improves results on ErdΕsβ minimum-overlap problem.
- The method demonstrates that simpler approaches can outperform complex evolutionary strategies in test-time program discovery.
- Token-level sampling is found to be a more effective source of diversity than random perturbations of model weights.
- The study establishes the largest evolution strategies training pipeline to date, showing that a zero learning rate can enhance maximum return.
Read more
Hill Sampling for Test-Time Scaling: A Simple and Better Alternative to Repeated Sampling, Evolution, and Training
Summary
This paper introduces Hill Sampling, a novel method for test-time program discovery using large language models (LLMs). The authors investigate the necessity of complex evolutionary search mechanisms and test-time training by proposing a simpler approach that retains the best candidate solution found so far and conditions subsequent samples on it. Hill Sampling is evaluated on three mathematical optimization problems: circle packing, sums/differences of sets, and ErdΕsβ minimum-overlap problem, achieving state-of-the-art results on circle packing and improving upon existing methods for the other two problems. The authors also conduct a large-scale study of evolution strategies applied to LLM weights, revealing that simpler methods outperform more complex ones, particularly emphasizing the effectiveness of token-level sampling over random model perturbations. The findings suggest that a straightforward sampling strategy can yield strong results without the need for additional complexity.
Methodology
The authors propose Hill Sampling, which involves repeatedly sampling candidate program edits from a frozen LLM, retaining the best program found, and conditioning subsequent samples on this program. They evaluate the method on three optimization problems using three different open-weight models and compare it against existing evolutionary strategies and sampling methods.
Results
Hill Sampling sets a new state-of-the-art result for circle packing in under five hours on eight NVIDIA H100 GPUs and improves upon the AlphaEvolve reference for ErdΕsβ minimum-overlap problem. The method shows strong performance on sums and differences of finite sets, with findings indicating that simpler methods yield better results than complex evolutionary strategies.
Implications
The findings suggest that researchers and practitioners can achieve effective program discovery with simpler methodologies, potentially reducing computational costs and complexity in the design of algorithms for LLMs. This could lead to more efficient applications in algorithmic discovery and optimization tasks.
You Only Need 2/3 of the Chosen Experts: An Empirical Study of Dynamic Expert Pruning in Fine-Grained MoE LLMs
Large Language Models
Efficient ML
NLP
- Fine-grained MoE architectures exhibit high redundancy in expert selection, allowing for significant pruning without loss of performance.
- Retaining two-thirds of selected experts maintains 98.8% of performance on average, leading to substantial speedups.
- Dynamic expert pruning methods can outperform uniform retention under aggressive pruning conditions, especially for generative tasks.
- Model sensitivity to pruning varies, with larger models being more robust compared to multimodal models.
Read more
You Only Need 2/3 of the Chosen Experts: An Empirical Study of Dynamic Expert Pruning in Fine-Grained MoE LLMs
Summary
This paper investigates the redundancy of expert selection in fine-grained mixture-of-experts (MoE) architectures for large language models (LLMs). The authors conduct a systematic empirical study across twelve MoE checkpoints from nine architecture families, utilizing a suite of eleven benchmarks that span knowledge question answering, mathematics, code generation, and general reasoning tasks. The study reveals that expert selection is significantly more redundant than previously assumed, with the finding that retaining approximately two-thirds of the selected experts preserves an average of 98.8% of performance compared to unpruned models. This approach yields a speedup of 1.2 to 1.7 times across two serving backends. The paper also evaluates dynamic expert pruning methods, showing that while simple uniform retention is effective, dynamic allocation can provide additional benefits under aggressive pruning conditions, particularly in generative tasks. The sensitivity to pruning varies by model type, with larger models exhibiting more resilience compared to multimodal models. Overall, the findings highlight the potential for reducing computational costs in fine-grained MoE models while informing future pruning strategies.
Methodology
The authors conducted a systematic empirical study involving twelve fine-grained MoE checkpoints across nine architecture families. They employed a core suite of eleven benchmarks to evaluate the performance of different expert selection strategies, including uniform retention and dynamic pruning methods. The study analyzed the redundancy of expert selection and the sensitivity of models to pruning based on various task types.
Results
The study found that uniformly retaining two-thirds of selected experts preserves an average of 98.8% of performance across models. This approach results in speedups of 1.2 to 1.7 times. Dynamic pruning methods showed improved performance over uniform retention under aggressive pruning scenarios, particularly in generative tasks, with gains of up to 3.0 points in performance.
Implications
The findings suggest that significant computational savings can be achieved in fine-grained MoE models without sacrificing performance. This has implications for the deployment of LLMs in resource-constrained environments and informs future research on expert pruning strategies, potentially leading to more efficient model architectures.
The Effect of Quantization on Clinical Benchmarks: Accuracy and Safety Across Model Families
NLP
Large Language Models
Efficient ML
- INT8 quantization is broadly safe for clinical deployment with minimal degradation.
- INT4 quantization leads to significant and model-dependent performance drops, especially in high-risk scenarios.
- Clinical fine-tuning does not guarantee robustness against quantization effects.
- Accuracy and safety robustness are independent properties that must be assessed separately.
Read more
The Effect of Quantization on Clinical Benchmarks: Accuracy and Safety Across Model Families
Summary
This paper investigates the impact of quantization on the clinical accuracy and safety of large language models (LLMs) when deployed on resource-constrained clinical edge devices. The authors evaluate five models with 7-8 billion parameters at different precision levels (FP16, GPTQ-INT8, and GPTQ-INT4) across five clinical benchmarks, including MedQA and MedSafetyBench. The study reveals that while INT8 quantization is generally safe with minimal degradation, INT4 quantization leads to significant performance drops that vary by model and task type. Notably, models fine-tuned for clinical tasks do not necessarily exhibit robustness against compression. The findings indicate that high-risk clinical scenarios are particularly vulnerable to performance degradation under INT4, and that safety and accuracy robustness are independent properties. The authors also explore recovery methods, finding that task-specific validation is essential for effective recovery strategies. Overall, the study emphasizes the need for careful evaluation of quantization effects on clinical models before deployment.
Methodology
The authors conducted a factorial evaluation of five 7-8B parameter models (two general-purpose and three clinical) across three quantization levels (FP16, GPTQ-INT8, GPTQ-INT4) and five clinical benchmarks. They measured performance degradation and safety alignment, particularly focusing on high-risk clinical tasks.
Results
The study found that INT8 GPTQ is safe across all models tested, with a maximum degradation of -1.9%. In contrast, INT4 quantization resulted in substantial degradation, with the BioMistral-7B model losing 19.7% accuracy in MedMCQA. High-risk scenarios showed even greater vulnerability, with Qwen2.5-7B experiencing a -26.8% drop under INT4. The results also indicated that safety degradation varied significantly across different models and tasks.
Implications
These findings highlight the importance of quantization evaluation in clinical settings, particularly for high-risk applications. The results suggest that while INT8 quantization is a viable option for deploying LLMs in healthcare, INT4 quantization poses risks that require careful consideration. The independence of accuracy and safety robustness implies that both must be prioritized in model development and deployment strategies.
An Exploratory Replica-Overlap Probe of the Grokking Transition
Theory
- The study aimed to explore the changes in pairwise weight overlaps during the grokking transition using RSB methods.
- The registered statistical tests did not validate, resulting in an undetermined outcome regarding the overlap distribution.
- No significant evidence was found for changes in the shape of the overlap distribution across the grokking transition.
- A post-hoc analysis indicated a descriptive relationship between grokking frequency and training fraction, though confounded by split identity.
Read more
An Exploratory Replica-Overlap Probe of the Grokking Transition
Summary
This paper investigates the phenomenon of 'grokking' in neural networks, where a model's accuracy suddenly improves after a period of training loss saturation. The authors trained 64 independently seeded networks across four configurations to explore whether the distribution of pairwise weight overlaps changes during the grokking transition, utilizing a replica-symmetry-breaking (RSB) approach inspired by statistical physics. Despite pre-registering their statistical tests, the study found that the registered implementation failed to validate the instrument used, leading to an 'undetermined' outcome. The analysis revealed no significant evidence that the shape of the overlap distribution changes across the grokking transition. Additionally, a post-hoc examination suggested a descriptive association between grokking frequency and training fraction, though this was confounded by split identity. The findings highlight the challenges in measuring the effects of grokking and suggest that the statistical methods employed may require refinement for future investigations.
Methodology
The authors trained 64 independently seeded neural networks in four configurations, analyzing pairwise weight overlaps using a replica-symmetry-breaking approach. They employed statistical tests to assess the distribution of overlaps and the validity of their measurements.
Results
The study did not yield a confirmatory null result due to instrument validity issues, and no significant changes in the overlap distribution were observed across the grokking transition. The only configuration that met the checkpoint-completeness requirement showed an uninformative result in the statistical analysis.
Implications
The results suggest that the methods for probing grokking in neural networks may need refinement. Understanding grokking could have implications for improving training strategies and model performance in machine learning.
Exploring Solver-Level Warmstarting for Neural Network Verification
Optimization
Theory
Efficient ML
- First study on solver-level warmstarting for neural network verification at the MILP solver level.
- Demonstrates significant reductions in verification time and improved solvability of instances.
- Investigates multiple warmstarting scenarios and their effects on different property changes.
- Presents a proof-of-concept pipeline using the SYMPHONY solver for empirical evaluation.
Read more
Exploring Solver-Level Warmstarting for Neural Network Verification
Summary
This paper investigates the application of solver-level warmstarting techniques to enhance the efficiency of neural network verification, which is crucial for ensuring the robustness of neural networks in safety-critical applications. The authors highlight that traditional verification methods treat each task independently, which can be inefficient given the similarities across various verification instances. By leveraging warmstarting, which reuses information from previous solver runs, the authors aim to reduce verification time and improve the ability to solve instances that are otherwise intractable. The study employs the SYMPHONY MILP solver to systematically analyze the impact of warmstarting across different property modifications, such as perturbation radii and input variations. The results demonstrate significant performance improvements, showcasing that warmstarting not only reduces running time but also enables the successful verification of previously unsolvable instances. This work represents a novel approach by focusing on solver-level warmstarting rather than verifier-level mechanisms, providing a proof-of-concept pipeline that can be adapted for state-of-the-art verifiers.
Methodology
The authors developed a pipeline that integrates warmstarting techniques into neural network verification processes using the SYMPHONY MILP solver. They systematically explored the effects of various property changes on the efficiency of warmstarting, analyzing multiple scenarios and strategies to assess their impact on verification performance.
Results
The experimental results indicate that solver-level warmstarting can lead to substantial improvements in verification times across various scenarios. Notably, warmstarting allowed for the successful verification of instances that could not be solved from scratch within the time limits, demonstrating its effectiveness in practical applications.
Implications
The findings suggest that incorporating solver-level warmstarting into neural network verification can significantly enhance the efficiency and effectiveness of verifying neural networks, particularly in safety-critical domains. This approach could be integrated into existing verification frameworks to improve their performance and reliability.
MICRO: Multi-Fidelity Active Search for Severe Error Discovery
NLP
Audio & Speech
Efficient ML
- MICRO effectively combines strong and weak feedback to maximize severe error discovery.
- The framework uses a joint Gaussian model to infer relationships between quality ratings and annotation losses.
- Experiments show that MICRO significantly outperforms existing policies in terms of severe error discovery.
- The approach allows for efficient allocation of a shared budget for feedback acquisition.
Read more
MICRO: Multi-Fidelity Active Search for Severe Error Discovery
Summary
The paper introduces MICRO (Multi-Fidelity Impact Clustered Rollout), an innovative active search framework designed to optimize the discovery of severe errors in machine learning models by leveraging both strong and weak human feedback under a shared budget. The authors highlight the challenge of varying costs and informativeness of human feedback, particularly in contexts like automatic speech recognition (ASR) and machine translation, where strong feedback (e.g., detailed corrections) is significantly more expensive than weak feedback (e.g., simple quality ratings). MICRO addresses this by jointly modeling annotation losses and quality ratings based on item features, allowing it to effectively allocate resources for feedback acquisition. The framework clusters potential acquisitions by their predicted impact on severity probabilities, ensuring a diverse selection of candidates for evaluation. Experimental results on the WMT20 English-German dataset demonstrate that MICRO outperforms several comparison policies, achieving the highest mean discovery count across different budget and rating cost scenarios, thus confirming its effectiveness in severe error discovery.
Methodology
MICRO employs a multi-fidelity active search strategy that integrates both strong and weak feedback types. It uses a joint Gaussian model to reconstruct feedback and estimate annotation losses based on item features. The framework clusters potential feedback candidates based on their predicted impact on severity probabilities and utilizes a rollout mechanism to evaluate their discovery value. This allows for informed decision-making under a constrained acquisition budget.
Results
The experimental evaluation on the WMT20 English-German dataset indicates that MICRO achieves the highest mean discovery count across various budget and rating cost settings. It shows significant improvements over six comparison policies, including two rollout controls, with statistical significance (p < .001) in three of the four tested scenarios.
Implications
The findings suggest that MICRO can enhance error discovery processes in machine learning applications, particularly in areas where human feedback is critical but costly. This framework could be beneficial in refining models in fields such as natural language processing and automatic speech recognition, where understanding and correcting severe errors is crucial for performance.
GLR-MM: Graph-Based Global-Local Reconstruction for Robust Multimodal Chest X-ray and EHR Representation Learning under Missing Modalities
Multimodal
Graph Learning
- GLR-MM effectively reconstructs missing modalities using both local and global patient data.
- The framework employs a graph-based approach to enhance prediction accuracy under missing data conditions.
- GLR-MM shows superior performance compared to existing methods, especially with high levels of missing modalities.
- The methodology includes adaptive fusion of reconstructed embeddings to optimize predictions.
Read more
GLR-MM: Graph-Based Global-Local Reconstruction for Robust Multimodal Chest X-ray and EHR Representation Learning under Missing Modalities
Summary
The paper introduces GLR-MM, a novel framework designed to enhance early ICU mortality prediction by effectively handling missing modalities in multimodal clinical data, specifically chest X-rays (CXR) and electronic health records (EHR). Traditional models often struggle with incomplete data, leading to suboptimal predictions. GLR-MM addresses this by employing a graph-based approach that reconstructs missing embeddings using both local (intra-patient) and global (inter-patient) evidence. The framework consists of two branches: a local branch that predicts missing data from the patient's available modalities and a global branch that retrieves information from similar patients within a graph structure. This dual approach allows for adaptive fusion of the reconstructed data, optimizing the prediction process. The authors conducted experiments on a dataset of 9,620 ICU stays with varying levels of missing modalities, demonstrating that GLR-MM outperforms existing methods, particularly in scenarios with significant missing data. The results indicate that leveraging graph-based reconstruction is particularly beneficial when a substantial portion of the input modalities is absent.
Methodology
GLR-MM utilizes a Graph-Based Global-Local Reconstruction framework that consists of two branches: a local branch for intra-patient reconstruction and a global branch for inter-patient reconstruction. It employs graph attention mechanisms to fuse reconstructed embeddings adaptively and optimizes for class-balanced prediction and reconstruction objectives. The model is evaluated on a dataset with controlled missingness levels, allowing for a comprehensive analysis of its performance.
Results
In experiments with 10%, 30%, and 50% random modality missingness, GLR-MM achieved higher AUROC and AUPRC scores compared to the baseline method MUSE, particularly excelling at 50% missingness with improvements of 0.0088 in AUROC and 0.0249 in AUPRC. These results underscore the effectiveness of graph-guided reconstruction in scenarios with significant data incompleteness.
Implications
The findings suggest that GLR-MM can significantly improve clinical decision-making processes in critical care settings where timely predictions are essential, even with incomplete data. This framework could be adapted for other multimodal clinical applications, enhancing the robustness of predictive models in healthcare.
RS-Claw-Evolution: Environment-Feedback-Driven Evolution for Lightweight Remote Sensing Agents in Long-Horizon Tasks
Reinforcement Learning
Large Language Models
NLP
- RS-Claw-Evolution framework improves lightweight RS agents' performance in long-horizon tasks.
- Three-stage evolution process: Interaction, Experience, and Decision Evolution.
- Feedback-driven strategies enhance learning from environmental interactions.
- Experimental results show significant performance improvements over baseline models.
Read more
RS-Claw-Evolution: Environment-Feedback-Driven Evolution for Lightweight Remote Sensing Agents in Long-Horizon Tasks
Summary
This paper introduces RS-Claw-Evolution, a novel framework designed to enhance the capabilities of lightweight remote sensing (RS) agents in executing long-horizon tasks. Traditional lightweight RS agents face significant challenges, including the loss of long-horizon states, inefficient use of environmental feedback, and sparse optimization signals. The proposed framework addresses these issues through a three-stage evolution process: (1) Interaction Evolution, which employs a programming-based paradigm to manage observations and maintain intermediate states, reducing context redundancy; (2) Experience Evolution, which utilizes a feedback-driven supervised fine-tuning strategy to generate informative error patterns and preserve recovery behaviors; and (3) Decision Evolution, which incorporates reinforcement learning to optimize tool-use behaviors and mitigate credit assignment problems. Experimental results on the Earth-Bench benchmark demonstrate that the optimized Qwen3-4B-based agent achieves an accuracy of 65.9% in Autonomous Planning mode, significantly outperforming the untrained Qwen3-32B baseline and approaching the performance of larger models like GPT-5. This indicates that effective integration of environmental feedback can significantly enhance the performance of lightweight agents in complex remote sensing tasks.
Methodology
The RS-Claw-Evolution framework is structured into three progressive stages: Interaction Evolution uses a programming-based approach to manage observations; Experience Evolution employs a feedback-driven supervised fine-tuning strategy to generate informative error patterns; and Decision Evolution utilizes reinforcement learning to optimize tool-use behaviors and address credit assignment issues.
Results
The optimized Qwen3-4B-based agent achieved an accuracy of 65.9% in Autonomous Planning mode, outperforming the untrained Qwen3-32B baseline (43.8%) and surpassing large-scale models like DeepSeek-V3.1 (60.8%), while nearing GPT-5's performance (71.6%).
Implications
The findings suggest that integrating environmental feedback into the training of lightweight RS agents can significantly enhance their performance, making them more competitive with larger models in complex geospatial analysis tasks. This has potential applications in automated Earth observation and remote sensing.
Minimal Recurrent Behavioral Memory for Imitation under Partial Observability
Robotics
Reinforcement Learning
Theory
- Introduces a framework for minimal recurrent memory in imitation tasks under partial observability.
- Characterizes memory requirements using conditional entropy and compatibility relations.
- Demonstrates that learned code rates remain near theoretical limits despite increasing hidden modes.
- Identifies challenges in learning representations and the impact of supervision on performance.
Read more
Minimal Recurrent Behavioral Memory for Imitation under Partial Observability
Summary
This paper addresses the challenge of determining the minimal recurrent memory required to imitate an expert's behavior under conditions of partial observability. The authors introduce a framework that characterizes this minimal memory through a compatibility relation, which allows for the preservation of behavioral distinctions that cannot be restored by future observations. They demonstrate that under transitive compatibility, the conditional entropy of the expert's behavioral quotient provides the exact minimum memory requirement. The study includes a measurement protocol that differentiates between behavioral sufficiency, excess code rate, and the information carried by observations. Experimental results show that learned code rates remain close to theoretical limits even as hidden modes increase, and that anticipatory memory requirements can be effectively managed. The paper also explores the challenges of learning this representation, noting that event-agnostic supervision can improve performance but may also introduce predictive surplus. Overall, the authors provide a novel representation target and validate it through rigorous experiments, highlighting the complexities of learning in environments with partial observability.
Methodology
The authors develop a theoretical framework based on conditional entropy and compatibility relations to characterize minimal recurrent memory. They employ a measurement protocol to assess behavioral sufficiency and code rates, and conduct experiments across various manipulation tasks to validate their findings. The methodology includes both theoretical analysis and empirical testing to ensure robustness.
Results
The results indicate that the minimal recurrent memory required for effective imitation under partial observability can be accurately characterized. The experiments show that learned code rates approach zero and two-bit requirements as the number of hidden modes increases, and that anticipatory memory can be managed effectively. The study also reveals that event-agnostic supervision can enhance learning but may lead to surplus information, which can be mitigated through annealing techniques.
Implications
The findings have significant implications for the design of learning algorithms in robotics and other fields where imitation learning is crucial. The ability to determine minimal memory requirements can lead to more efficient learning systems that can operate effectively under partial observability, enhancing the performance of autonomous agents in complex environments.
Optimal No-Regret Learning for Repeated Prophet Inequality
Theory
Optimization
- Introduces an efficient algorithm for repeated prophet inequalities achieving eO(βT) expected regret.
- Eliminates polynomial dependence on the number of boxes, addressing a significant limitation in previous works.
- Combines empirical backward induction with box-specific reach bonuses to enhance exploration strategies.
- Resolves an open question in the field regarding optimal regret bounds under prefix feedback.
Read more
Optimal No-Regret Learning for Repeated Prophet Inequality
Summary
This paper addresses the problem of repeated prophet inequalities under prefix feedback, where a learner must make sequential decisions about accepting values drawn from unknown distributions. The author presents an efficient algorithm that achieves an expected regret of eO(βT), which matches the lower bound up to logarithmic factors. The algorithm explores near-optimal policies by combining empirical backward induction with box-specific reach bonuses, thus eliminating the polynomial dependence on the number of boxes. This work resolves an open question posed by Liu et al. (2025) regarding the simultaneous achievement of optimal regret bounds without dependence on the number of boxes. The study highlights the challenges of exploration in the context of sequential decision-making and proposes a novel approach that balances exploration and exploitation effectively.
Methodology
The methodology involves maintaining a behavior policy with O(Ξ΅) suboptimality and using empirical backward induction to construct box-specific explorers. The algorithm leverages the nesting structure of observed prefixes to optimize exploration and combines these strategies with a relative-drop aggregation rule to minimize regret.
Results
The proposed algorithm achieves a regret bound of O(βT [1 + log(nT)]Β²), which matches the β¦(βT) lower bound up to logarithmic factors. This represents a significant improvement over previous results, providing both optimal horizon dependence and reduced dependence on the number of boxes.
Implications
The findings have implications for sequential decision-making in various applications, including online learning, resource allocation, and adaptive systems where decisions must be made under uncertainty. The approach could enhance strategies in fields requiring efficient exploration of unknown distributions.
Block-Level Weight-Space Structure Persists Under Post-Training: An Empirical Study Across LLM Families
Large Language Models
- Post-training modifies all tensors but preserves block-level weight-space structure.
- Mean cosine similarity across blocks exceeds 0.99, indicating high structural retention.
- Independently trained models show lower similarity, indicating disconnected weight-space regions.
- LinkerLLM achieves 18-48% GPU memory savings by sharing blocks across variants.
Read more
Block-Level Weight-Space Structure Persists Under Post-Training: An Empirical Study Across LLM Families
Summary
This paper presents an empirical study on the effects of post-training on the weight-space geometry of large language models (LLMs). The study investigates eight configurations across four architecture families (Qwen2.5, Llama-3.1/3.2, Mistral, Gemma-2) and identifies a significant phenomenon termed the 'granularity gap.' While post-training modifies every tensor in the models, it remarkably preserves the block-level structure of the weights. The authors report that all tensors change, with no identical tensors remaining post-training, yet the mean cosine similarity between blocks exceeds 0.99 and the relative Frobenius distance remains below 0.13. This indicates that post-training acts as a structured perturbation, shifting parameters while maintaining block-level geometry. The study also reveals that independently trained specializations show a lower similarity to the general base model, suggesting a disconnected region in weight space. The perturbation magnitude varies with model scale, architecture family, and post-training methodology. As a practical application, the authors introduce LinkerLLM, a lazy loader that allows for sharing of blocks across co-resident variants, achieving significant GPU memory savings. The results demonstrate that five out of eight configurations retain over 94% of the quality of the unshared variant on various benchmarks, while the remaining configurations show minor drops in performance.
Methodology
The study employs two metrics to assess the similarity and perturbation of transformer blocks between base and post-trained models: cosine similarity and Frobenius distance. The analysis covers eight base-to-instruct pairs across four architecture families, comparing weights in float32 after conversion from their native precision.
Results
The analysis reveals that only 2 out of 339 tensors in the Qwen2.5-7B model are byte-identical post-training, while no tensors match in Llama-3.1-8B. Despite this, block-level similarity remains high, with cosine similarities above 0.99 and Frobenius ratios below 0.05 for most families. The perturbation magnitude varies based on model scale and training methodology, with significant implications for model performance.
Implications
The findings suggest that post-training can be leveraged to optimize memory usage in LLM deployments without sacrificing performance. The introduction of LinkerLLM could facilitate more efficient model serving and resource allocation in practical applications.
What Should a Self-Teacher See? Privileged Context Design for On-Policy Self-Distillation
Large Language Models
NLP
Theory
- Concise semantic contexts can outperform full solutions in self-distillation tasks.
- Answer-only context remains competitive, especially at larger model scales.
- Initial teacher-student KL divergence does not correlate with teaching effectiveness.
- The preferred context varies based on the student model's scale and the task at hand.
Read more
What Should a Self-Teacher See? Privileged Context Design for On-Policy Self-Distillation
Summary
This paper investigates the design of privileged contexts in on-policy self-distillation (OPSD) for machine learning models, particularly focusing on how the level of abstraction in the context provided to a self-teacher affects the learning outcomes of the student model. The authors argue that more privileged information does not necessarily lead to better teaching performance. They compare the conventional full solution context with three alternative abstractions: a named strategy, a method-independent framing, and a problem category, as well as an answer-only control. The study is conducted on competition mathematics tasks, revealing that intermediate contexts can outperform full solutions while requiring significantly fewer hint tokens. The results indicate that the optimal context varies depending on the student model's scale and the specific task, suggesting that a self-teacher should provide context that aligns with the student's capacity to act on it. The findings challenge the assumption that more detailed contexts always enhance learning, emphasizing the importance of context design in self-distillation frameworks.
Methodology
The authors employed controlled interventions to assess the impact of different context types on a fixed self-teacher in OPSD. They constructed three progressively abstract context packages from a problem-solution pair using an offline compiler and compared these with a full solution, an answer-only condition, and a control. The study involved evaluating student performance across multiple scales and benchmarks while maintaining fixed parameters for the student prompt, teacher checkpoint, and training set.
Results
The study found that the best intermediate contexts improved in-domain peak means by 1.4 points at 4B and 1.6 points at 8B compared to the full solution, while using significantly fewer hint tokens. The answer-only condition was competitive, trailing the full solution by only 0.2 points at larger scales. Notably, the initial KL divergence between teacher and student did not predict downstream performance, indicating that context design is crucial for effective self-distillation.
Implications
The findings suggest that in designing self-distillation frameworks, practitioners should focus on the level of abstraction in the context provided to the self-teacher, as it can significantly influence the learning outcomes. This has potential applications in optimizing training processes for various machine learning models, particularly in scenarios where computational resources are limited.
Can We Predict Anomaly Detection Performance from Embedding-Space Geometry?
Audio & Speech
Theory
- Anomaly detection performance can be predicted using embedding-space geometry without anomalous data.
- A lower bound on AUC is derived, linking performance to inlier and outlier score variances and separation.
- Inlier score variance alone is insufficient for reliable model selection; pseudo-anomaly probes improve performance estimation.
- Experiments show that pseudo-anomaly-based estimators outperform conventional selection methods under domain shifts.
Read more
Can We Predict Anomaly Detection Performance from Embedding-Space Geometry?
Summary
This paper investigates the feasibility of predicting the performance of anomaly detection systems without access to anomalous data, focusing on k-nearest neighbor (kNN) based detectors. The authors derive a lower bound on the area under the ROC curve (AUC) that connects detection performance to the separation between inlier and outlier scores, as well as their variances. They analyze how factors such as density variation, intrinsic-dimensional heterogeneity, and cross-domain mismatch affect score variability. The study reveals that inlier score variance alone is not a reliable predictor of performance across different representations. To overcome this limitation, the authors introduce pseudo-anomaly probes, which serve as references for estimating score separation. Experiments conducted on the DCASE 2022β2025 benchmarks demonstrate that these pseudo-anomaly-based estimators significantly enhance anomaly-free model selection, outperforming traditional methods, especially under domain shifts. The findings underscore the predictive power of embedding-space geometry regarding anomaly detection performance and highlight the representation-dependent nature of performance estimates derived from inlier-only data.
Methodology
The authors derive a theoretical framework based on the geometry of the embedding space, specifically focusing on kNN-based detectors. They analyze the relationship between score variances and performance metrics, and introduce pseudo-anomaly probes to estimate score separation. The methodology includes experiments on various embedding models and datasets to validate their findings.
Results
The introduction of pseudo-anomaly probes significantly improved the accuracy of anomaly-free model selection, allowing it to outperform traditional methods that rely on labeled anomalies. The experiments demonstrated that diverse pseudo-anomalies could effectively enhance performance estimation, particularly in scenarios involving domain shifts.
Implications
The findings suggest that anomaly detection systems can be optimized without the need for labeled anomalies, making them more applicable in real-world scenarios where anomalies are rare or difficult to obtain. This could lead to more robust and efficient anomaly detection methods across various domains, including audio and speech processing.
Provably Efficient Reinforcement Learning in Continuous-Time Episodic MDPs with Poisson Decision Epochs
Reinforcement Learning
Theory
Efficient ML
- Introduces a continuous-time framework for episodic MDPs with Poisson decision epochs.
- Extends UCRL and Q-learning algorithms to achieve O(T^(2/3)) regret bounds.
- Establishes matching minimax lower bounds, confirming the optimality of the proposed rates.
- Addresses the exploration-exploitation tradeoff in continuous domains with Lipschitz continuity.
Read more
Provably Efficient Reinforcement Learning in Continuous-Time Episodic MDPs with Poisson Decision Epochs
Summary
This paper addresses the challenges of reinforcement learning (RL) in continuous-time episodic Markov Decision Processes (MDPs) where decision epochs are determined by a homogeneous Poisson process. The authors introduce a framework that allows for smooth variations in reward and transition dynamics over time, which is crucial for applications in robotics, recommendation systems, and resource allocation. They extend existing RL algorithms, specifically UCRL and Q-learning, to this continuous-time setting under a Lipschitz continuity assumption. The paper proves that both model-based and model-free algorithms achieve a regret bound of O(T^(2/3)), establishing the first tight regret guarantees for Lipschitz-smooth continuous-time episodic MDPs with Poisson decision epochs. Additionally, the authors provide matching minimax lower bounds, demonstrating that their results are optimal up to logarithmic factors. This work significantly advances the theoretical understanding of RL in continuous domains and opens avenues for practical applications where decision-making occurs at irregular intervals.
Methodology
The authors formalize a continuous-time episodic MDP framework and utilize Lipschitz continuity to exploit local smoothness. They adapt existing RL algorithms (UCRL and Q-learning) through discretization and confidence sets tailored for the continuous setting. The analysis includes controlling randomness in decision epochs and deriving high-probability bounds to ensure statistical efficiency.
Results
The paper demonstrates that both the model-based and model-free algorithms achieve a regret bound of O(T^(2/3)). Furthermore, it establishes matching minimax lower bounds of β¦(T^(2/3)), confirming that the proposed regret rates are optimal up to logarithmic factors.
Implications
The findings have significant implications for real-world applications of RL in continuous domains, such as robotics and resource allocation, where decisions are made at irregular intervals. The theoretical guarantees provided can enhance the reliability and efficiency of RL algorithms in these contexts.
An Exact Junction-Tree Extended Formulation for Optimal Classification Trees
Optimization
Theory
Interpretability
- Introduces an exact LP formulation for bounded-depth classification trees using junction-tree representation.
- Supports recursive subtree optimization and employs exact reductions to minimize model size.
- Presents two solution methods: column generation and message passing for efficient computation.
- Demonstrates significant runtime improvements and optimality certification over existing methods.
Read more
An Exact Junction-Tree Extended Formulation for Optimal Classification Trees
Summary
This paper presents an exact linear programming (LP) formulation for bounded-depth classification trees utilizing a junction-tree representation. The proposed formulation is integral and facilitates recursive subtree optimization. By employing exact reductions, the model's size is significantly minimized while maintaining the optimal value and enabling the recovery of an optimal tree. The authors introduce two solution methods: column generation and message passing. Column generation addresses integral restricted LPs and certifies optimality using bounds over the full feasible domain, while message passing combines optimal subtree costs recursively. The computational experiments demonstrate that the exact reductions lead to a substantial decrease in the size of the junction-tree formulation. The resulting LP formulation is capable of certifying instances that the tested mixed-integer formulation fails to establish optimality for within the same computational budget. Furthermore, the column-generation and message-passing methods are shown to certify more instances and achieve a significant reduction in geometric-mean runtime compared to existing state-of-the-art methods for optimal classification trees.
Methodology
The authors develop an exact LP formulation that leverages junction-tree theory to represent dependencies among splits and observations in classification trees. The formulation allows for recursive optimization of subtrees, which can be combined to form a complete tree. Two main solution techniques are employed: column generation, which solves integral restricted LPs, and message passing, which recursively combines subtree costs.
Results
The computational experiments reveal that the proposed exact reductions lead to a notable decrease in the size of the junction-tree formulation. The new LP formulation successfully certifies optimality for instances where traditional mixed-integer formulations do not, achieving an order-of-magnitude reduction in geometric-mean runtime compared to existing methods.
Implications
The findings suggest that the proposed junction-tree formulation can enhance the efficiency and effectiveness of optimal classification tree construction, making it a valuable tool for applications requiring interpretable decision-making and predictive performance. This approach could be particularly beneficial in fields such as finance, healthcare, and any domain where decision trees are commonly used for classification tasks.
Efficient Mixture-of-Experts with Speculative Decoding via Expert Coactivation
Large Language Models
Efficient ML
NLP
- Combining MoE models with SD can significantly improve inference speed but poses challenges related to memory transfer costs.
- High expert coactivation during training leads to faster runtimes and improved throughput.
- A combination of training techniques enhances expert coactivation without altering the SD algorithm.
- The proposed method achieves a 21% improvement in throughput over standard MoE models while retaining accuracy.
Read more
Efficient Mixture-of-Experts with Speculative Decoding via Expert Coactivation
Summary
This paper addresses the challenges of combining Mixture-of-Experts (MoE) models with Speculative Decoding (SD) to enhance inference speed while managing memory transfer costs. The authors identify that the efficiency of SD with MoEs is significantly influenced by the number of tokens verified, which in turn affects the number of experts transferred from DRAM to the Neural Processing Unit (NPU). The study reveals that routers designed for high expert coactivation can lead to improved runtime performance. By implementing a series of training-time modificationsβincluding a global load-balancing loss, shared experts, a consistency loss, and an autoregressive expert selection mechanismβthe authors achieve enhanced expert coactivation. This results in a 21% increase in throughput compared to standard MoE baselines, while maintaining comparable accuracy. The findings underscore the importance of router design in optimizing MoE performance during inference with SD.
Methodology
The authors propose a novel approach to improve expert coactivation in MoE models during training. They replace the standard local load-balancing loss with a global one, utilize shared experts, introduce a consistency loss to align neighboring token selections, and implement an autoregressive expert selection mechanism. These modifications aim to reduce the number of unique experts transferred to the NPU during SD, thereby enhancing throughput.
Results
The proposed method results in a 21% increase in throughput compared to baseline MoE models while maintaining similar accuracy levels. The study demonstrates that improved expert coactivation leads to reduced memory transfer costs and enhanced efficiency in MoE models during inference.
Implications
The findings suggest that optimizing router design and training methodologies can significantly enhance the performance of MoE models in real-world applications, particularly in scenarios requiring rapid inference, such as natural language processing tasks and large-scale language models.
Belted Engression: Sufficient Dimension Reduction for Generative Distributional Regression
Generative Models
Theory
Efficient ML
- Introduces Belted Engression, a new framework for generative distributional regression.
- Establishes a compress-then-generate paradigm that integrates sufficient dimension reduction into generative models.
- Proves theoretical equivalence between SDR conditions and law-preserving generative factorization.
- Demonstrates superior performance in distributional prediction with fewer parameters compared to existing methods.
Read more
Belted Engression: Sufficient Dimension Reduction for Generative Distributional Regression
Summary
The paper introduces Belted Engression, a novel framework for generative distributional regression that addresses the challenges of learning complex covariate dependencies in modern conditional generative models. Traditional sufficient dimension reduction (SDR) methods have not been tailored for conditional generation, leading to inefficiencies in handling high-dimensional data. Belted Engression proposes a compress-then-generate paradigm that integrates SDR into the generative architecture, allowing for a low-dimensional representation of covariates before stochastic generation. The authors establish a theoretical foundation proving that the SDR condition aligns with a law-preserving generative factorization at the global optimum of the Belted Engression objective. They also derive sharper finite-sample convergence rates and demonstrate that the belted architecture operates with fewer parameters than traditional methods. Extensive simulations and real-world applications show that Belted Engression outperforms existing methods in distributional prediction and SDR recovery, making it a significant advancement in the field of generative models.
Methodology
The methodology involves a unified energy-score objective that jointly learns a sufficient predictor mapping and a reduced stochastic generator. This approach embeds a structural bottleneck into the generative architecture, allowing for efficient representation learning and conditional generation.
Results
The results indicate that Belted Engression achieves better distributional prediction and SDR recovery compared to traditional methods, with a significantly reduced number of trainable parameters. The theoretical results also confirm the efficiency and effectiveness of the proposed framework.
Implications
The implications of this research extend to various fields requiring accurate modeling of complex conditional distributions, such as environmental science and biology. The framework can enhance predictive modeling capabilities and improve the interpretability of generative models in high-dimensional settings.
Matched-Input Estimates Differ in Sign Across Architectures: Auditing EEG Foundation Models on Motor Imagery
Time Series
- Pretrained EEG foundation models show inconsistent performance transfer to downstream tasks.
- A validation-locked evaluation protocol was developed to prevent data leakage.
- Supervised models consistently outperform foundation models in motor imagery tasks.
- Matched-input analysis reveals accuracy differences that vary in sign across architectures.
Read more
Matched-Input Estimates Differ in Sign Across Architectures: Auditing EEG Foundation Models on Motor Imagery
Summary
This paper investigates the performance of pretrained EEG foundation models, specifically LaBraM and CBraMod, in the context of motor imagery tasks for brain-computer interfaces (BCIs). The authors highlight inconsistencies in the transferability of representations from these models to downstream tasks, particularly in light of different evaluation protocols and input pipelines. They employ a validation-locked evaluation protocol to ensure that all tunable decisions are made using training-session data only, thereby preventing data leakage. The study reveals that every supervised comparator outperforms the foundation models on the BCI Competition IV-2a dataset, indicating a significant performance gap. Furthermore, the authors conduct a matched-input analysis across three supervised architectures, finding that the accuracy differences vary in sign, suggesting that a single comparator may not adequately capture the performance gap across architectures. The results also indicate that the observed performance deficit does not uniformly replicate across different motor-imagery datasets, and that temperature scaling can effectively calibrate foundation model outputs despite lower accuracy. Overall, this work emphasizes the need for standardized evaluation protocols and highlights the complexities involved in comparing foundation models with task-specific decoders.
Methodology
The authors implemented a validation-locked evaluation protocol where all tunable decisions were made using training-session data only. They conducted a matched-input analysis by retraining three supervised architectures on both broadband and narrowband EEG recordings to assess the impact of input pipeline differences on model performance.
Results
The study found that every supervised model outperformed the foundation models on the BCI Competition IV-2a dataset. The matched-input analysis showed that accuracy differences varied in sign across architectures, with broadband input improving ATCNet accuracy while reducing EEG Conformer accuracy. Additionally, the performance gap did not replicate uniformly across different datasets, and temperature scaling effectively calibrated foundation model outputs.
Implications
The findings suggest that EEG foundation models may not be universally applicable across different tasks and highlight the importance of standardized evaluation protocols in the field of brain-computer interfaces. The results also indicate potential avenues for improving model performance through careful consideration of input pipelines and calibration techniques.
A Distributional Optimisation Perspective on Combining Models in Deep Learning
Optimization
Theory
Large Language Models
- Introduces a distributional optimisation framework for model combination in deep learning.
- Reformulates ensemble methods and LoRA averaging as entropy-regularised optimisation problems.
- Demonstrates that the ensemble case has a convex objective, while the LoRA averaging case does not.
- Evaluates various algorithms, including a new variant of variational gradient descent.
Read more
A Distributional Optimisation Perspective on Combining Models in Deep Learning
Summary
This paper explores the potential of distributional optimisation in the context of combining models in deep learning, aiming to improve the performance of machine learning tasks. The authors identify that traditional methods for combining models, such as ensemble methods and low-rank adapter (LoRA) averaging, often rely on heuristic approaches for training individual models and their combination strategies. They propose a principled framework that treats the collection of models as a discrete distribution, optimising this distribution jointly rather than independently. The paper makes three key contributions: (1) it reformulates ensemble and LoRA averaging strategies as entropy-regularised distributional optimisation problems, revealing that the ensemble case has a convex objective while the LoRA averaging does not; (2) it evaluates existing and new algorithms for this optimisation task, including a functional variant of variational gradient descent (VGD); and (3) it presents empirical results from synthetic classification tasks and fine-tuning large language models on commonsense reasoning benchmarks, demonstrating the effectiveness of the proposed methods. The findings suggest that a distributional approach can lead to more efficient model training and better predictive performance by systematically considering the interplay between model training and combination.
Methodology
The authors cast model combination strategies as optimisation tasks over probability distributions. They formulate loss functions for ensembles and LoRA averaging, and explore algorithms that can effectively optimise these distributions. The methodology includes both theoretical analysis and empirical evaluation on synthetic and real-world tasks.
Results
The empirical studies indicate that the proposed distributional optimisation framework can enhance the performance of combined models, outperforming traditional heuristic methods. The results highlight the benefits of joint training and systematic optimisation in improving predictive accuracy and robustness.
Implications
This work suggests that adopting a distributional optimisation perspective can lead to more efficient training of deep learning models, potentially reducing computational waste and improving model performance. It opens avenues for further research into principled model combination strategies in various machine learning applications.
Beyond Imitation: Auditing the Recoverability of Reasoning in Distilled Models
NLP
Large Language Models
Theory
- Prefix recovery measures the ability of student models to complete reasoning tasks based on partial teacher solutions.
- Reverse-KL distillation shows consistent improvements in performance, especially for smaller models.
- As student model size increases, prefix recovery rates improve significantly, indicating better reasoning compatibility.
- The study identifies a capacity-dependent transfer regime where the benefits of distillation decrease with larger models.
Read more
Beyond Imitation: Auditing the Recoverability of Reasoning in Distilled Models
Summary
This paper investigates the compatibility of reasoning transfer from teacher to student models in knowledge distillation, focusing on the concept of prefix recovery. The authors propose a method to measure how well a student model can complete a reasoning task after being given a portion of a verified solution. They explore the relationship between prefix recovery and the cosine conflict between cross-entropy and reverse-KL gradients. The study finds that reverse-KL distillation significantly improves performance, particularly for smaller student models, and that as model size increases, the ability to recover reasoning improves. The authors also identify a capacity-dependent transfer regime where the effectiveness of distillation diminishes as models become larger and more capable. The findings suggest that prefix recovery can serve as a diagnostic tool for selecting appropriate distillation strategies.
Methodology
The authors conducted experiments using adjacent teacher-student model pairs with varying parameter sizes. They measured prefix recovery by revealing portions of verified solutions and assessing student completions. They also analyzed the cosine conflict between cross-entropy and reverse-KL gradients to understand the compatibility of reasoning transfer.
Results
The results indicate that prefix recovery improves from 71.0% to 91.9% as student model size increases from 0.6B to 4B parameters. The robust-fragile recovery gap shrinks significantly, and the objective conflict decreases as well, suggesting better alignment between teacher and student models. Additionally, reverse-KL distillation enhances accuracy on benchmark tasks, with notable improvements for smaller models.
Implications
The findings have implications for optimizing knowledge distillation strategies, particularly in selecting teacher models based on their reasoning compatibility with student models. The concept of prefix recovery can guide future research in model training and evaluation, especially in the context of reasoning tasks.
Towards Adaptive Federated Graph Clustering: A Global Community-aware Contrastive Learning-based Approach
Graph Learning
Federated Learning
- Identifies critical limitations in existing federated graph clustering methods.
- Proposes AdaFGC, which uses global community anchors for adaptive clustering.
- Implements a community-aware contrastive learning scheme to enhance clustering performance.
- Demonstrates state-of-the-art performance across multiple datasets and metrics.
Read more
Towards Adaptive Federated Graph Clustering: A Global Community-aware Contrastive Learning-based Approach
Summary
This paper addresses the challenges of federated graph clustering (FGC), which allows multiple clients to collaboratively train graph models without sharing private data. The authors identify two critical limitations in existing FGC methods: unrealistic pre-defined cluster cardinality and incomplete inter-community separation. To overcome these challenges, they propose AdaFGC, an adaptive framework that utilizes global community-aware contrastive learning. AdaFGC introduces an over-complete set of global community anchors to model community structures and adaptively estimate clustering cardinality through cross-client anchor refinement. The framework employs a contrastive learning scheme that enforces community-level attraction and repulsion while stabilizing local representations through node-level and topology-level objectives. Extensive experiments on eight benchmark datasets show that AdaFGC consistently outperforms existing supervised and unsupervised FGL baselines across multiple clustering metrics, demonstrating its effectiveness in addressing the complexities of federated graph clustering.
Methodology
The methodology involves initializing an over-complete set of global community anchors from local cluster centroids, followed by an adaptive refinement process where clients assign nodes to the nearest anchors and the server evolves these anchors based on clustering results. The global community-aware contrastive learning scheme optimizes community-level, node-level, and topology-level objectives to improve representation learning.
Results
AdaFGC was tested on eight benchmark datasets, showing consistent improvements over existing supervised and unsupervised federated graph learning methods across four clustering metrics, indicating its robustness and effectiveness in real-world applications.
Implications
The proposed AdaFGC framework has significant implications for unsupervised learning in federated settings, particularly in scenarios where data privacy is crucial, and labeled data is scarce. It can be applied in various domains, such as social network analysis, recommendation systems, and bioinformatics, where distributed graph data is prevalent.
Gap-Free Streaming PCA Beyond Rank-One Updates: Near-Optimal Rates and Applications to Differential Privacy
Theory
Efficient ML
- Introduces a gap-free variant of streaming PCA using Oja's algorithm.
- Achieves near-optimal convergence rates without eigengap assumptions.
- Extends results to energy PCA and differentially private PCA.
- Utilizes a second moment bound for stochastic updates, simplifying previous methodologies.
Read more
Gap-Free Streaming PCA Beyond Rank-One Updates: Near-Optimal Rates and Applications to Differential Privacy
Summary
This paper presents a new analysis of Oja's algorithm for streaming principal component analysis (PCA) without the need for eigengap assumptions, addressing a significant gap in existing literature. The authors demonstrate that Oja's algorithm can achieve near-optimal convergence rates under weaker regularity conditions than previously required. They introduce a correlation PCA (cPCA) objective that allows for the recovery of leading eigenvectors even when eigenvalues are closely spaced. The paper also extends the results to energy PCA (ePCA) and provides guarantees for differentially private PCA in a gap-free setting. The methodology relies on a second moment bound for stochastic updates, which simplifies previous approaches that required stronger assumptions. The results include sample complexity bounds that are competitive with existing methods, and the authors provide a nearly-matching lower bound for their convergence guarantees. Additionally, they settle an open question regarding differentially private PCA, confirming that their approach can achieve similar rates as those under eigengap assumptions, thus broadening the applicability of PCA in privacy-sensitive contexts.
Methodology
The authors analyze Oja's algorithm under a general framework that does not require eigengap assumptions. They employ a second moment bound on stochastic updates to derive convergence guarantees. The analysis includes extensions to correlation PCA and energy PCA, as well as applications to differentially private PCA, leveraging geometric aggregation techniques for improved sample complexity.
Results
The main results include a convergence guarantee for Oja's algorithm that holds with constant probability, alongside a nearly-matching lower bound. The authors also provide sample complexity bounds for both correlation PCA and energy PCA, demonstrating that their methods can achieve similar performance to existing approaches under eigengap assumptions. For differentially private PCA, they establish sample bounds that are competitive with prior work, confirming the effectiveness of their approach in privacy-sensitive applications.
Implications
The findings have significant implications for the field of machine learning, particularly in scenarios where data privacy is a concern. The ability to perform PCA without eigengap assumptions expands the applicability of these techniques in real-world data streams, making it easier to extract meaningful insights while adhering to privacy constraints. This work also opens avenues for further research in streaming algorithms and differential privacy.
EMGBlend: Heterogeneity-Aware Self-Supervised Pretraining for Gesture and Force Decoding
Time Series
- EMGBlend addresses the challenges of heterogeneous EMG datasets through a self-supervised learning framework.
- The framework employs shared channel patches, geometry-aware attention, and source-exposure balancing.
- Pretraining on 11 public EMG sources leads to improved performance in gesture recognition and force decoding.
- Ablation studies validate the importance of the proposed mechanisms in enhancing transfer learning.
Read more
EMGBlend: Heterogeneity-Aware Self-Supervised Pretraining for Gesture and Force Decoding
Summary
The paper presents EMGBlend, a self-supervised framework designed to address the challenges posed by the heterogeneity of public surface electromyography (EMG) datasets. These datasets vary significantly in terms of electrode layout, channel count, frequency support, and size, which can lead to misalignment and dominance of certain datasets during training. EMGBlend introduces a novel approach that combines shared channel patches with geometry-aware attention, restricts spectral targets to the frequency bands supported by each recording, and balances exposure across different data sources. The authors pretrain a 109M-parameter model using 11 public EMG sources and evaluate its performance on tasks such as gesture recognition, continuous-force regression, and contact classification. The results demonstrate that EMGBlend consistently outperforms models initialized randomly and those based on waveform reconstruction. Furthermore, the study shows that multi-source pretraining enhances gesture recognition and maintains competitive performance for force decoding, while ablation studies confirm the contributions of the proposed mechanisms. Overall, EMGBlend effectively combines heterogeneous EMG datasets through a carefully designed mechanism rather than simple concatenation.
Methodology
The methodology involves pooling heterogeneous EMG sources while addressing differences in channel counts, sampling rates, and frequency support. EMGBlend utilizes shared channel patches and geometry-aware attention to create a common interface for variable sensor layouts. It restricts spectral targets to valid frequency bands for each recording and employs source-exposure correction to balance the influence of different datasets. The model is pretrained using masked representation learning and vector quantization techniques.
Results
The pretrained model outperformed matched random initialization and waveform reconstruction controls across various tasks, including gesture recognition, continuous-force regression, and contact classification. Multi-source pretraining improved gesture recognition performance and remained competitive for force decoding tasks. Ablation studies confirmed that the design choices significantly contributed to the model's transfer capabilities.
Implications
The findings suggest that EMGBlend can enhance the performance of EMG-based applications in gesture recognition and force estimation by effectively leveraging diverse datasets. This approach may lead to better generalization in real-world applications where sensor configurations vary.
BOBA: Dynamic Bayesian Optimization through Bayesian Active Inference
Optimization
Robotics
Time Series
- BOBA is a novel acquisition function for dynamic Bayesian optimization that minimizes predictive uncertainty.
- It incorporates principles from active inference to enhance exploration-exploitation trade-offs.
- Experiments show that BOBA significantly reduces regret in query-restricted settings compared to existing DBO methods.
- The paper provides insights into tuning exploration strategies for different types of dynamic functions.
Read more
BOBA: Dynamic Bayesian Optimization through Bayesian Active Inference
Summary
The paper introduces BOBA (Bayesian Optimization through Bayesian Active Inference), a novel acquisition function designed to address the challenges of dynamic black-box optimization (DBO) where the objective function changes over time. Traditional DBO methods often rely on standard acquisition functions like Upper Confidence Bound (UCB), which do not adequately account for temporal variations, leading to inefficient sample allocation and poor tracking of moving optima. BOBA leverages principles from active inference to minimize predictive uncertainty about future states in dynamic environments. This approach enhances the exploration-exploitation trade-off by incorporating a forward-looking uncertainty quantification that estimates potential changes in the objective function. The authors evaluate BOBA against state-of-the-art DBO methods on synthetic dynamic benchmarks, demonstrating significant improvements in regret in query-restricted scenarios while maintaining competitiveness in time-limited settings. The paper also explores various exploration strategies within BOBA, providing insights into tuning the exploration-exploitation balance for different dynamic functions. Overall, BOBA represents a significant advancement in DBO methodologies, offering both a new acquisition function and a framework for real-time optimization in non-stationary environments.
Methodology
The authors propose BOBA as an acquisition function that integrates free energy principles from active inference to optimize sample selection in dynamic environments. They evaluate BOBA against existing DBO methods using synthetic benchmarks, analyzing its performance in terms of regret and exploration-exploitation balance.
Results
The evaluation of BOBA demonstrates that it significantly improves regret in query-restricted settings compared to state-of-the-art DBO methods. Additionally, it remains competitive in scenarios with time constraints. Variants of BOBA with different exploration strategies further highlight the flexibility and effectiveness of the proposed method.
Implications
The development of BOBA has significant implications for real-time optimization tasks in various fields, including robotics, automated systems, and any application requiring continuous adaptation to changing environments. It offers a new approach to handle dynamic optimization problems more effectively.
Entropy Can Flow, or It Can Guide. Be Entropy. LEDFlow: Introducing Entropy-guided Generation Order into Uniform Discrete Flow
Generative Models
Theory
Multimodal
- LEDFLOW reduces the rate of correct-to-wrong revisions in generated outputs.
- The method employs entropy-guided absorption to prioritize reliable predictions.
- LEDFLOW achieves state-of-the-art performance in Sudoku solving and multimodal tasks.
- Theoretical analysis confirms the effectiveness of local entropy in guiding absorption decisions.
Read more
Entropy Can Flow, or It Can Guide. Be Entropy. LEDFlow: Introducing Entropy-guided Generation Order into Uniform Discrete Flow
Summary
This paper introduces Low-Entropy Discrete Flow (LEDFLOW), a novel approach to improve the generation order in uniform discrete flow by incorporating entropy-guided absorption. The authors highlight the issue of correct intermediate predictions being overwritten by incorrect final outputs, particularly in tasks like Sudoku, where 9.4% of generated cells were found to be correct at some point but incorrect in the final output. LEDFLOW addresses this by selectively fixing reliable predictions while allowing others to evolve, thus minimizing the risk of absorbing incorrect predictions. The method employs a binary absorption indicator to manage the absorption process based on local predictive entropy, ensuring that the most reliable predictions are fixed first. Theoretical analysis supports the approach by demonstrating that selecting low-entropy positions minimizes conditional prediction errors. Experimental results show that LEDFLOW significantly reduces the rate of incorrect final outputs and achieves high accuracy across various reasoning tasks, including a 0.845 accuracy in Sudoku solving and superior performance in text-to-image generation and multimodal understanding tasks.
Methodology
LEDFLOW introduces a training-free sampling method that uses local predictive entropy to guide the absorption of predictions in uniform discrete flow. The method involves a binary absorption indicator that allows reliable predictions to be fixed while other positions continue to evolve. The approach is theoretically justified through a Kullback-Leibler decomposition of absorption error, analyzing joint dependence and conditional prediction terms.
Results
LEDFLOW achieved a Sudoku solve accuracy of 0.845 and outperformed native samplers in text-to-image generation and multimodal understanding across six benchmarks. The method reduced the rate of incorrect final outputs from 9.4% to 2.6% and significantly decreased the number of token changes occurring late in the sampling process.
Implications
The findings suggest that entropy-guided methods can enhance the reliability of generative models in various applications, particularly in structured reasoning tasks and multimodal generation. This could lead to more robust AI systems capable of handling complex decision-making scenarios.