AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
24
Papers today
8h
Update frequency
7
Days of history
Observation-Grounded Self-Predictive Reinforcement Learning for Visual Continuous Control
Reinforcement Learning
Computer Vision
Robotics
- Introduces OG-SPR, a model-free RL algorithm that combines self-prediction and observation prediction for improved data efficiency.
- Utilizes lightweight adapters to mitigate over-constraining effects of auxiliary objectives on shared representations.
- Demonstrates significant performance improvements on 28 visual control tasks compared to state-of-the-art methods.
- Addresses the limitations of existing self-predictive and observation-predictive methods in challenging visual control domains.
Read more
Observation-Grounded Self-Predictive Reinforcement Learning for Visual Continuous Control
Summary
This paper addresses the challenge of sample-efficient policy learning from pixel inputs in reinforcement learning (RL), particularly in visual continuous control tasks. The authors propose a novel algorithm called Observation-Grounded Self-Predictive Representations (OG-SPR), which integrates two auxiliary objectives: multi-step latent self-prediction and next-observation prediction. The motivation behind OG-SPR is to leverage the complementary strengths of self-predictive and observation-predictive methods, which have shown promise in improving data efficiency but have limitations when used in isolation. The authors argue that self-prediction focuses on temporal predictability in latent space, while observation prediction grounds representations in observation-level dynamics. To mitigate the over-constraining effect of directly combining these objectives, OG-SPR introduces lightweight adapters that allow the shared representation to benefit from predictive signals without being forced to satisfy the self-prediction objective directly. The algorithm is evaluated on 28 visual control tasks from the DeepMind Control Suite, demonstrating significant performance improvements over existing state-of-the-art methods, particularly in challenging environments such as dog and humanoid tasks. Overall, OG-SPR represents a promising step towards enhancing data efficiency in model-free visual RL by effectively integrating different predictive learning strategies.
Methodology
OG-SPR is built on an off-policy actor-critic framework and incorporates two core auxiliary tasks: multi-step latent self-prediction and next-observation prediction. Additionally, it employs short-term value prediction to stabilize value learning. The algorithm introduces lightweight adapters for latent self-prediction to allow the shared representation to receive predictive signals without direct imposition of the self-prediction objective.
Results
Experimental results show that OG-SPR outperforms state-of-the-art self-predictive and observation-predictive RL methods across 28 visual control tasks, achieving particularly strong gains in complex environments such as dog and humanoid tasks.
Implications
The findings suggest that integrating different predictive learning strategies can significantly enhance the data efficiency of model-free visual RL, paving the way for future research in this area. This could have applications in robotics and other domains requiring efficient learning from visual inputs.
An Emerging Retail Portfolio Management Application: Personalized, Tax-Aware Reinforcement Learning with Natural Language Goals
Reinforcement Learning
NLP
Optimization
- Development of a personalized, tax-aware portfolio management application for retail investors.
- Utilization of a three-phase reinforcement learning system for portfolio recommendations.
- Integration with a live brokerage API for real-time portfolio management.
- Preliminary validation through backtesting, with practical engineering insights shared.
Read more
An Emerging Retail Portfolio Management Application: Personalized, Tax-Aware Reinforcement Learning with Natural Language Goals
Summary
This paper addresses the gap in personalized, tax-aware portfolio management for retail investors, which is typically available only to institutional clients. The author presents a fully integrated application that allows users to articulate investment goals in natural language, which are then mapped to specific investment mandates. The application utilizes a three-phase reinforcement learning system comprising a self-supervised cross-asset encoder, a Mixture-of-Experts (MoE) allocation policy, and a lightweight LoRA adapter for personalization based on individual brokerage behavior. The system has been integration-tested with a live brokerage API (Alpaca) and includes features such as multi-user authentication and an auditable action-integrity chain. The paper discusses practical engineering lessons learned during development and presents preliminary validation through 14-day walk-forward backtests, although the application has not yet been opened to real end-users. The findings highlight the potential for sophisticated portfolio management tools to be accessible to individual investors, bridging the gap between institutional and retail investment strategies.
Methodology
The methodology involves a three-phase reinforcement learning approach: Phase 1 focuses on cross-asset representation learning using a self-supervised encoder; Phase 2 employs a Mixture-of-Experts policy for portfolio allocation; and Phase 3 personalizes recommendations using a lightweight LoRA adapter based on user behavior without retraining the model.
Results
The application has been successfully integration-tested with a live brokerage API, demonstrating the ability to generate personalized portfolio recommendations based on user-defined goals. Preliminary backtesting results indicate the system's potential effectiveness, although it has not yet been deployed for real trading.
Implications
The development of this application could democratize access to sophisticated portfolio management tools for retail investors, enabling them to make informed investment decisions that consider tax implications and personal financial goals. It also highlights the potential for using reinforcement learning and natural language processing in financial technology applications.
QEvict: Recoverable Quantized KV Eviction for Attention-Drift-Robust Long-Context Decoding
NLP
Large Language Models
Efficient ML
- QEvict introduces a recoverable eviction strategy for KV caches, allowing previously evicted tokens to be reinstated based on changing importance.
- The method employs a three-tier cache management system that balances full precision and quantization to optimize memory usage.
- QEvict significantly reduces missed attention and enhances information retention in long-context decoding tasks.
- The approach is validated against existing methods, showing consistent performance improvements across various benchmarks.
Read more
QEvict: Recoverable Quantized KV Eviction for Attention-Drift-Robust Long-Context Decoding
Summary
The paper introduces QEvict, a novel three-tier Key-Value (KV) cache management system designed to enhance the efficiency of autoregressive large language model inference, particularly in long-context settings. Traditional eviction methods in KV caches often make irreversible decisions that can lead to the loss of important tokens, especially as the importance of tokens can change dynamically during decoding. QEvict addresses this limitation by implementing a recoverable eviction strategy that allows for the promotion of previously evicted tokens back to full precision when their importance increases. The system categorizes windows of tokens into three tiers: high-confidence windows retained in full precision, intermediate windows stored in a quantized recoverable format, and low-confidence windows that are evicted. This dynamic approach not only preserves historical context but also reduces missed attention and improves information retention across various long-context tasks. The authors validate QEvict against existing eviction and quantization methods, demonstrating significant improvements in performance metrics across benchmarks for understanding, retrieval, and reasoning tasks.
Methodology
The authors developed QEvict, a three-tier KV-cache management scheme that categorizes token windows into high-confidence, quantized recoverable, and low-confidence tiers. They introduced diagnostics like Future Missed Mass and Global LIR to assess the importance of tokens dynamically. The system updates cumulative attention scores to manage the promotion and demotion of token windows, allowing for a flexible and recoverable eviction process.
Results
QEvict consistently outperformed traditional eviction and quantization baselines across long-context understanding, retrieval, and reasoning benchmarks, leading to reduced missed attention and improved information retention.
Implications
The findings suggest that QEvict can significantly enhance the performance of large language models in applications requiring long-context processing, such as document summarization, question answering, and complex reasoning tasks, by improving memory management and token importance handling.
A Unified Risk View of Uncertainty: Posterior Risk for Disentanglement and Evaluation Beyond Proxies
Theory
Interpretability
- Introduces a unified definition of uncertainty as pointwise posterior risk.
- Develops a benchmark for evaluating uncertainty estimates without relying on proxy tasks.
- Demonstrates that accurate predictions do not guarantee reliable uncertainty estimates.
- Identifies meaningful differences between methods in terms of their alignment with oracle uncertainty.
Read more
A Unified Risk View of Uncertainty: Posterior Risk for Disentanglement and Evaluation Beyond Proxies
Summary
This paper addresses the critical need for reliable uncertainty estimates in safety-sensitive applications, emphasizing the importance of disentangling epistemic and aleatoric uncertainty. The authors highlight the inconsistencies in the definitions of these uncertainty types across the literature, which complicates the evaluation of uncertainty estimation methods. They propose a unified definition of uncertainty as pointwise posterior risk, which represents the expected loss of a predictor under the distribution of plausible ground-truth functions given the data. This approach integrates Bayesian uncertainty with estimator-dependent deviations, thus providing a comprehensive framework for uncertainty evaluation. The authors introduce a theory-backed benchmark that allows for the direct computation of oracle epistemic and aleatoric uncertainty using semi-synthetic datasets with known generative processes. This benchmark facilitates a detailed analysis of uncertainty estimates, revealing that accurate predictions do not necessarily correlate with reliable uncertainty disentanglement. The findings indicate significant differences between various methods, highlighting their alignment with oracle uncertainty targets and their sensitivity to dataset and modeling choices.
Methodology
The authors propose a unified definition of uncertainty and develop a benchmark that utilizes semi-synthetic datasets with real covariates and known generative processes. This allows for the direct computation of oracle epistemic and aleatoric uncertainty, facilitating a more accurate evaluation of uncertainty estimation methods.
Results
The empirical analysis reveals that methods with high predictive accuracy do not always provide reliable uncertainty estimates. The benchmark effectively distinguishes between methods, identifying those that align closely with oracle uncertainty targets while also exposing their sensitivity to different datasets and modeling choices.
Implications
The findings of this paper have significant implications for the deployment of AI systems in safety-critical domains, where reliable uncertainty estimates are essential for decision-making. The proposed benchmark can guide practitioners in selecting appropriate uncertainty estimation methods and improve the interpretability and safety of AI applications.
How Far Do Simple Transformations Translate Across Text Embedding Models?
NLP
Theory
Efficient ML
- The study systematically evaluates simple translation methods across diverse text embedding models.
- Compatibility of transformations is highly pair-dependent, influenced by architectural and training factors.
- Simple transformations can recover shared structures but are not universally applicable across models.
- The research highlights the limitations of existing literature that suggests universal latent compatibility.
Read more
How Far Do Simple Transformations Translate Across Text Embedding Models?
Summary
This paper investigates the effectiveness of simple transformations, specifically linear mappings, in translating representations across various text embedding models. The authors aim to understand the compatibility of independently trained models in organizing semantic information, which is crucial for enabling AI-to-AI communication without the need for human-readable text. The study evaluates nine different embedding models that vary in architecture, pooling strategies, and training objectives. Using multiple diagnostic methods, including Centered Kernel Alignment (CKA), downstream transfer, fidelity, and retrieval, the authors find that while simple transformations can recover meaningful structures and facilitate transfer between some compatible model pairs, they fail for others. The compatibility of transformations is shown to depend on a combination of factors such as model architecture, training objectives, pooling methods, and data distributions. The findings challenge the notion of universal compatibility among embedding spaces, suggesting that simple mappings are not as broadly applicable as previously thought.
Methodology
The authors conducted a systematic empirical study involving nine text embedding models with varying characteristics. They employed linear mappings as simple translators and evaluated their effectiveness using multiple diagnostics, including CKA for geometric similarity, downstream transfer performance, fidelity, and retrieval capabilities. The study included both constrained linear maps based on small anchor sets and more flexible linear maps trained on larger paired corpora.
Results
The results indicate that while simple transformations can effectively recover meaningful shared structures between some model pairs, they fail to do so for others. The degree of compatibility is influenced by the specific architecture, training objectives, pooling strategies, and data distributions of the models involved. Overall, the study provides evidence against the notion of universal compatibility among heterogeneous embedding spaces.
Implications
The findings suggest that while simple transformations can facilitate some level of interoperability between text embedding models, the limitations identified may impact the development of modular AI systems and agent-to-agent communication frameworks. Understanding these limitations is crucial for future research in AI communication and the design of embedding-based tools.
Neuro-Symbolic Closed-Loop Control of Laser Powder Bed Fusion with an In-Loop Ontology
Optimization
Robotics
Theory
- Introduces a neuro-symbolic architecture for closed-loop control in LPBF.
- Integrates an ontology for real-time symbolic reasoning with statistical learning.
- Addresses the challenge of geometry-dependent melt pool quality, specifically overhang dross.
- Demonstrates superior performance in maintaining quality compared to traditional methods.
Read more
Neuro-Symbolic Closed-Loop Control of Laser Powder Bed Fusion with an In-Loop Ontology
Summary
This paper presents a novel neuro-symbolic closed-loop control architecture for laser powder bed fusion (LPBF) that integrates an ontology within the control loop. The proposed system addresses the challenge of geometry-dependent melt pool quality in LPBF by combining symbolic reasoning with statistical learning. The ontology links process objectives and constraints to observable signals, allowing for real-time adjustments to the control parameters. Specifically, the architecture focuses on managing overhang dross, a defect caused by excessive melt pool depth that cannot be directly measured during the build process. By mapping the geometry and power-dependent depth-to-width ratio onto observable width constraints, the system effectively maintains quality while adapting to changes in geometric context. The architecture demonstrates superior performance compared to traditional geometry-blind controllers, achieving zero dross with minimal residual defects and maintaining robustness under varying conditions. The results highlight the feasibility of this approach, paving the way for future experimental calibrations and applications in additive manufacturing.
Methodology
The methodology involves a geometry-conditioned neuro-symbolic closed-loop control architecture where an ontology operates within the control loop. This ontology translates geometric and material information into references and constraints for a model predictive controller (MPC). A description-logic reasoner dynamically selects active constraints based on the geometric context, while a Gaussian process provides calibrated uncertainty for safety margins.
Results
The proposed architecture successfully eliminates dross produced by a geometry-blind controller, maintaining zero dross levels with only minor residual lack of fusion. It demonstrates graceful degradation under plant mismatches and can retarget to new alloys and constraints by simply editing ontology data, rather than requiring code changes.
Implications
The findings suggest significant advancements in the control of additive manufacturing processes, particularly in enhancing the quality and reliability of LPBF. The neuro-symbolic approach could lead to more adaptive and intelligent manufacturing systems capable of handling complex geometries and varying materials.
Continual Learning in Transition
NLP
Large Language Models
Reinforcement Learning
- Continual learning is transitioning from parameter-centric methods to system-level adaptations.
- The authors introduce a tri-axial framework to analyze CL evolution across dimensions of timing, mechanism, and capability locus.
- Emerging paradigms like on-policy learning and test-time training are reshaping CL methodologies.
- External components such as memory and skill libraries are becoming crucial for enhancing model adaptability.
Read more
Continual Learning in Transition
Summary
The paper discusses the evolution of continual learning (CL) from a parameter-centric approach to a more system-level adaptation framework. Traditional CL methods have focused on retaining knowledge through parameter updates, but recent advancements suggest a shift towards broader mechanisms that include on-policy learning, test-time training, and external components like memory and skill libraries. The authors propose a tri-axial framework to characterize this transition, examining the dimensions of 'When', 'How', and 'Where' learning occurs. They highlight the need for continual adaptation throughout the model lifecycle, from pre-training to inference, and emphasize the importance of external harness components in enhancing model capabilities. The paper systematically surveys existing methods, identifies key challenges, and discusses future directions for research in CL, particularly in the context of large language models and agentic AI, which require adaptive systems capable of evolving in dynamic environments.
Methodology
The authors systematically survey existing continual learning methods, categorizing them based on their learning mechanisms (off-policy, on-policy, beyond-gradient), timing (pre-training, post-training, inference), and capability locus (internal parameters vs. external harness components). They analyze the implications of these categories on the evolution of CL and propose a framework to understand the ongoing transition.
Results
The paper highlights the limitations of classical CL approaches in the face of rapid advancements in AI and proposes a more comprehensive understanding of CL that includes external mechanisms and adaptive strategies. It identifies the need for continual adaptation throughout the model lifecycle and discusses the implications of emerging paradigms for future research.
Implications
The findings suggest that future AI systems, particularly those aimed at achieving AGI, must incorporate continual learning strategies that allow for dynamic adaptation and knowledge accumulation in real-time. This has potential applications in various fields, including robotics, natural language processing, and interactive AI systems.
Kastor: An efficient fine-tuning strategy for generative emulation of PDE simulations
Generative Models
Efficient ML
Time Series
- Kastor introduces a two-stage inference scheme to enhance accuracy and efficiency in PDE simulations.
- Mean Prediction Regularization (MPR) significantly improves the stability and performance of generative models.
- Incorporating spatial gradient matching enhances the accuracy and physical fidelity of simulations.
- Kastor outperforms existing methods in forecasting accuracy and computational efficiency across multiple datasets.
Read more
Kastor: An efficient fine-tuning strategy for generative emulation of PDE simulations
Summary
The paper presents Kastor, a novel methodology designed to enhance the efficiency and accuracy of generative emulation for Partial Differential Equation (PDE) simulations. Traditional PDE solvers are computationally intensive, and while machine learning offers a potential solution, existing auto-regressive models often struggle with error accumulation and fail to capture the stochastic nature of complex systems. Kastor addresses these challenges through a two-stage inference scheme that combines a large-stride causal auto-regressive model with a non-causal temporal super-resolution network, significantly reducing error accumulation while minimizing computational costs. Additionally, the authors introduce Mean Prediction Regularization (MPR), a training objective that constrains the generative model to predict the deterministic distribution mean under null noise conditions, enhancing the performance and stability of both Functional Generative Networks (FGN) and diffusion-based emulators. The incorporation of spatial gradient matching further improves accuracy and physical fidelity, as evidenced by power spectrum density measurements. The methodology is evaluated using diverse datasets from The Well benchmark, demonstrating superior performance in forecasting accuracy, spectral consistency, and computational efficiency compared to existing methods, achieving a 42.9% average reduction in forecasting error over the reference Walrus fine-tuning methodology.
Methodology
The methodology involves a two-stage inference scheme combining a causal auto-regressive model with a temporal super-resolution network. It employs Mean Prediction Regularization (MPR) to constrain predictions and uses spatial gradient matching to enhance accuracy. The model is trained on diverse simulation datasets from The Well benchmark.
Results
Kastor demonstrates a 42.9% average reduction in forecasting error compared to the Walrus fine-tuning methodology, outperforming it in 8 out of 10 datasets on variance-normalized RMSE (VRMSE). The model shows improved forecasting accuracy, spectral consistency, and computational efficiency.
Implications
The advancements presented in Kastor could lead to more efficient and accurate simulations in various scientific and engineering applications, enabling faster design iterations and better understanding of complex physical phenomena.
When Do Corrective Features Help? An Agent for Corrective Feature Discovery on Black-Box Forecasters
Time Series
Interpretability
- CRAFTER introduces a source-blind framework for corrective feature discovery, focusing on residual correction rather than direct model improvement.
- The methodology combines compositional searches and LLM-generated features to enhance forecasting accuracy.
- CRAFTER outperforms existing feature-engineering systems across six datasets and various frozen backbone models.
- The effectiveness of corrective features is regime-dependent, with significant improvements observed in weak backbones.
Read more
When Do Corrective Features Help? An Agent for Corrective Feature Discovery on Black-Box Forecasters
Summary
This paper addresses the challenge of improving the performance of frozen pretrained forecasters, which often exhibit systematic errors that are costly to correct through fine-tuning. The authors propose a novel approach called CRAFTER (Corrective Residual Agent with Feature-based Temporal Exploration and Reasoning) that focuses on corrective feature discovery. Unlike traditional automated feature engineering that models the data-generating process, CRAFTER targets the model-failure process by mining interpretable features from the residuals of the frozen forecaster. The methodology employs two generators: a compositional search over raw input channels and a large language model (LLM) that suggests named feature combinations and executable code. A validation-grounded gate evaluates candidate features without bias towards their source, allowing for a fair comparison with other feature-engineering systems. The results demonstrate that CRAFTER consistently outperforms existing feature-engineering methods across multiple datasets and backbone models, significantly reducing forecasting errors. The study also characterizes the conditions under which corrective features are beneficial, inert, or harmful, providing insights into the effective use of corrective features in various forecasting scenarios.
Methodology
CRAFTER employs a dual-generator approach for corrective feature discovery: a compositional search over raw input channels and an LLM that proposes named feature combinations and executable code. A validation-grounded gate evaluates candidates without bias towards their source, and a validation-selected corrector applies the accepted features or retains the original forecast.
Results
CRAFTER consistently outperformed three dedicated feature-engineering systems across six public datasets and six frozen backbone families. The method achieved up to a 27% reduction in error for the weakest backbones and demonstrated robustness across different LLM backends and when fine-tuned.
Implications
The findings suggest that CRAFTER can serve as a deployable correction layer for frozen forecasters, enabling businesses to improve forecasting accuracy without the need for extensive retraining. The characterization of corrective features also aids in understanding the conditions under which these features can be effectively utilized.
LLM Inference Under Bursty Workload Distribution: Modifying the WAIT Algorithm
Large Language Models
Optimization
Efficient ML
- Proposes a modification to the WAIT scheduling algorithm for LLM inference.
- Introduces an online arrival rate estimation mechanism to adapt to bursty traffic.
- Utilizes a two-state Markov Modulated Poisson Process for modeling request patterns.
- Demonstrates improved throughput over existing algorithms in low arrival-rate scenarios.
Read more
LLM Inference Under Bursty Workload Distribution: Modifying the WAIT Algorithm
Summary
This paper addresses the challenges of scheduling algorithms for Large Language Model (LLM) inference under bursty workload distributions, which are common in real-world applications. The authors critique existing approaches that assume constant request arrival rates, highlighting their inadequacy in handling the dynamic nature of actual traffic. To overcome this limitation, they propose a lightweight modification to the WAIT scheduling algorithm that incorporates an online estimation of request intensity based on observed interarrival times. This modified algorithm utilizes a two-state Markov Modulated Poisson Process (MMPP-2) to model the temporal variations in request patterns. Through extensive simulation evaluations, the authors demonstrate that their modified WAIT algorithm achieves higher throughput compared to existing systems like Sarathi-Serve, ORCA, and vLLM in scenarios characterized by low arrival-rate shifts, while maintaining comparable latency. The findings suggest that the proposed method can effectively enhance the performance of LLM-serving systems under bursty workloads, making it a significant contribution to the field of LLM inference optimization.
Methodology
The authors developed a modified version of the WAIT algorithm that includes an online estimation mechanism for request intensity. They modeled the workload using a two-state Markov Modulated Poisson Process (MMPP-2) to reflect realistic traffic patterns. A comprehensive simulation framework was employed to evaluate the performance of the modified WAIT algorithm against other state-of-the-art systems under varying traffic conditions.
Results
The modified WAIT algorithm achieved higher throughput than Sarathi-Serve, ORCA, and vLLM in low arrival-rate shift scenarios, while maintaining similar latency levels. The simulation results indicate that the proposed method effectively handles bursty workloads, demonstrating its potential for improving LLM inference performance.
Implications
The findings of this research have significant implications for the deployment of LLMs in real-world applications, particularly in environments with bursty traffic patterns. The proposed scheduling algorithm can enhance the efficiency of LLM-serving systems, leading to improved user experiences in various domains such as customer support, education, and healthcare.
THBKG: A Temporal Biomedical Knowledge Graph for Decision-Aligned Clinical Advancement Prediction
Graph Learning
Time Series
Interpretability
- THBKG allows for the reconstruction of evidence profiles at specific past decision points.
- Graph propagation methods outperform traditional direct-evidence approaches in predicting clinical advancement.
- The model is particularly effective for target-disease pairs with no direct evidence at the time of decision-making.
- A path-based explainer enhances the interpretability of predictions by detailing the evidence landscape.
Read more
THBKG: A Temporal Biomedical Knowledge Graph for Decision-Aligned Clinical Advancement Prediction
Summary
The paper introduces the Temporal Heterogeneous Biomedical Knowledge Graph (THBKG), designed to enhance the prediction of clinical advancement for therapeutic target-disease pairs. The authors highlight that inadequate target-disease linkage is a significant contributor to Phase II efficacy failures in drug development. THBKG comprises 110,396 entities and 11.1 million edges across nineteen relation types, with each edge timestamped to reflect the year its evidence changed. This allows for the reconstruction of the evidence profile at the time of clinical decisions. The authors propose a decision-aligned benchmark for predicting whether a target-disease pair entering Phase II will advance to Phase III, based solely on evidence available at the time of the decision. The results demonstrate that graph propagation over THBKG significantly outperforms direct-evidence references, particularly for pairs lacking direct evidence at their decision point. The paper also presents a path-based explainer that provides insights into the evidence landscape behind predictions, promoting explainability in clinical advancement predictions. THBKG is released as a continually updated resource for retrospective validation of therapeutic target hypotheses.
Methodology
The authors developed the THBKG by integrating diverse biomedical data sources into a temporal knowledge graph, enabling the tracking of evidence changes over time. They defined a decision-aligned benchmark for predicting clinical advancement and employed graph propagation techniques to assess the predictive power of the knowledge graph. A ridge regression model was used to identify the top target-disease pairs based on engineered evidence features.
Results
The study found that graph propagation over THBKG achieved a relative success rate of 4.3β4.5 for the top ten pairs per therapeutic area, significantly outperforming direct-evidence references. Notably, the model was able to rank pairs five- to sixfold above chance, especially for the 72.8% of pairs lacking direct evidence at their decision point.
Implications
The THBKG can facilitate more accurate predictions of clinical advancement, potentially reducing the high costs associated with drug development failures. It provides a framework for retrospective validation of therapeutic hypotheses, which could enhance the efficiency of target identification and drug repurposing efforts.
Beyond Rotations: AuroOFT for Expressive Quantized Orthogonal Fine-Tuning
NLP
Large Language Models
Efficient ML
- AuroOFT enhances QOFT by introducing a nonlinear residual branch, improving expressivity without sacrificing stability.
- The method maintains orthogonality as a property of the QOFT branch while allowing for input-dependent corrections.
- AuroOFT demonstrates significant performance improvements over both QOFT and QLoRA in low-bit language model settings.
- The approach provides a framework for better utilizing parameter budgets in low-rank adaptations.
Read more
Beyond Rotations: AuroOFT for Expressive Quantized Orthogonal Fine-Tuning
Summary
The paper introduces AuroOFT, a novel approach to quantized orthogonal fine-tuning (QOFT) that enhances the expressivity of low-bit language models. While QOFT allows for parameter-efficient adaptation through structured activation rotations, it is limited to linear orthogonal transformations, which can restrict the model's ability to make input-dependent nonlinear corrections. AuroOFT addresses this limitation by adding a zero-start gated low-rank nonlinear residual to each adapted linear layer, allowing for more expressive corrections while maintaining the stability of the QOFT framework. The method maps activations into a compact latent space and employs adaptive nonlinear bases with bounded or token-dependent gating. AuroOFT is functionally equivalent to QOFT at initialization, preserving orthogonality as a branch-level property. Experimental results demonstrate that AuroOFT outperforms matched QOFT by 1.30β2.70 points in Macro-6 metrics on the 1.5B/3B Qwen2.5 settings and exceeds QLoRA by 6.52β10.62 points while reducing trainable parameters by 32.3β44.7% relative to QLoRA. The paper emphasizes the importance of balancing quantized stability with expressive correction capacity, particularly in mathematical reasoning tasks.
Methodology
AuroOFT employs a dual-branch architecture where one branch retains the QOFT structure for stable quantized adaptation, while the other introduces a zero-start gated nonlinear residual. This setup allows for the integration of nonlinear corrections without altering the original quantized weights, thus preserving the geometric properties of the model.
Results
AuroOFT achieved improvements of 1.30β2.70 points over matched QOFT and 6.52β10.62 points over QLoRA in Macro-6 metrics, while also reducing the number of trainable parameters by 32.3β44.7% compared to QLoRA.
Implications
The findings suggest that incorporating nonlinear corrections into quantized orthogonal frameworks can significantly enhance model performance, particularly in tasks requiring nuanced reasoning. This approach may lead to more efficient adaptations of large language models in resource-constrained environments.
Is Self-Pretraining really useful to improve diagnosis in medical Time Series?
Time Series
- SPT improves classification accuracy in medical time series tasks by 0-6 percentage points.
- The benefits of SPT are more pronounced in deeper transformer models.
- SPT can be applied effectively to both multivariate and univariate medical time series.
- No task-specific architectural changes are required to implement SPT.
Read more
Is Self-Pretraining really useful to improve diagnosis in medical Time Series?
Summary
This paper investigates the effectiveness of Self-PreTraining (SPT) in enhancing the performance of transformer architectures on medical time series data. The authors aim to determine if SPT can improve classification accuracy across various medical applications, particularly in scenarios with limited data. They evaluate transformer models on three medical time-series tasks: rehabilitation robotics, stress detection, and Parkinson's disease detection. The models are trained from scratch or via SPT using four different masking strategies to encourage temporal and cross-modal representation learning. The results indicate that SPT consistently enhances classification accuracy by 0-6 percentage points, with deeper models benefiting more from the enriched representations learned during pre-training. The findings suggest that SPT is a straightforward and effective method to boost transformer performance in medical time-series tasks without necessitating task-specific architectural modifications, highlighting its potential for improving robustness and accuracy in data-scarce clinical environments.
Methodology
The authors employed transformer architectures trained either from scratch or through Self-PreTraining (SPT) on three medical time series tasks. They utilized four masking-based objectives to promote learning of temporal and cross-modal representations. The study systematically varied model depth to analyze the interaction between model capacity and pre-training benefits.
Results
The study found that SPT consistently improved classification accuracy across all datasets and configurations, with gains ranging from 0 to 6 percentage points. The improvements were particularly significant for deeper models, which could better leverage the enriched representations obtained during pre-training.
Implications
The findings suggest that SPT can be a valuable strategy for enhancing the performance of machine learning models in medical diagnostics, particularly in situations where labeled data is scarce. This approach could lead to more robust and accurate diagnostic tools in clinical settings, ultimately improving patient outcomes.
Does Latent Context Help? A Controlled Evaluation of Inverse Reinforcement Learning in Arctic Shipping
Reinforcement Learning
Robotics
Interpretability
- Latent context in IRL may not provide additional benefits in real-world scenarios as previously assumed.
- A nonlinear shared reward model significantly outperforms a linear model in predicting vessel behavior.
- Observable route and environmental conditions explain most behavioral variations, questioning the necessity of latent context.
- The study introduces a context-need diagnostic for evaluating the relevance of latent context in decision-making.
Read more
Does Latent Context Help? A Controlled Evaluation of Inverse Reinforcement Learning in Arctic Shipping
Summary
This paper investigates the effectiveness of latent context in Inverse Reinforcement Learning (IRL) for Arctic shipping navigation, where reliable AI-assisted navigation is crucial due to rapidly changing sea-ice conditions. The authors conduct a controlled evaluation using 3,186 Automatic Identification System (AIS) derived voyages from 202 vessels over nine Arctic shipping seasons. They compare three reward models: a linear shared reward model (MCE-IRL), a nonlinear shared reward model (AIRL), and a latent-context model (PEMIRL) that incorporates vessel-specific latent context. The study finds that while the nonlinear reward model significantly outperforms the linear model by 50.9% in held-out likelihood, the addition of vessel-specific latent context actually reduces performance by 16.5%. Behavioral analysis reveals that variations in vessel behavior can largely be explained by observable factors such as route and environmental conditions, rather than hidden preferences. The authors emphasize the need for careful evaluation of latent context in AI systems, especially in safety-critical applications, and propose a context-need diagnostic to assess when latent context is beneficial. Overall, the findings challenge the assumption that latent context is necessary for capturing behavioral variation in real-world scenarios.
Methodology
The authors conducted a controlled evaluation using a dataset of 3,186 voyages from 202 vessels, comparing three reward models: MCE-IRL (linear shared reward), AIRL (nonlinear shared reward), and PEMIRL (latent-context model). They analyzed the performance of these models based on held-out likelihood, predictive accuracy, route fidelity, and reward transfer, while also employing behavioral analysis and context probes to understand the sources of behavioral variation.
Results
The nonlinear reward model (AIRL) improved held-out likelihood by 50.9% over the linear model (MCE-IRL). However, the introduction of vessel-specific latent context (PEMIRL) resulted in a 16.5% decrease in performance. The analysis indicated that observable factors largely accounted for behavioral variations, suggesting that latent context may not be necessary.
Implications
The findings suggest that AI systems in safety-critical domains, such as maritime navigation, should prioritize observable features over latent context to enhance reliability and interpretability. This could lead to more effective AI deployment in dynamic environments.
Spectral Distillation: From Nonlinear Dynamics to Linear State-Space Models
Theory
Robotics
Time Series
- Introduces a nonlinear-to-LDS distillation pipeline that separates learning from representation.
- Provides a dimension-free prediction guarantee based on observer complexity.
- Delivers an explicit compact representation of the learned system, suitable for long-horizon simulation and control.
- Empirical validation shows the distilled LDS can outperform traditional training methods.
Read more
Spectral Distillation: From Nonlinear Dynamics to Linear State-Space Models
Summary
This paper addresses the challenge of learning nonlinear dynamical systems using a compact linear state-space representation without directly solving a non-convex system identification problem. The authors propose a two-stage pipeline that begins with the learning of an implicit spectral predictor through Observation Spectral Filtering (OSF), a convex method that competes with the best linear observer for the system. The second stage involves converting this predictor into an explicit recurrent linear dynamical system (LDS) through a process called spectral-to-LDS distillation. The main theorem presented in the paper guarantees that the average prediction error of the distilled LDS can be decomposed into an exponentially small distillation term and an OSF learning term, which is governed by the observer complexity rather than the latent dimension of the nonlinear system. This approach yields the first end-to-end provable method for extracting a best-in-hindsight LDS representation from nonlinear dynamics. Empirical results demonstrate that the proposed train-then-distill pipeline produces compact LDS predictors that either match or outperform directly trained baselines across various benchmarks, including linear LDS and MuJoCo behavior cloning.
Methodology
The methodology consists of two main stages: first, learning an implicit spectral predictor using Observation Spectral Filtering (OSF), which operates in a fixed temporal feature space and does not attempt to recover latent states. Second, the learned spectral predictor is distilled into a finite-dimensional linear dynamical system (LDS) using recent results on spectral representations, ensuring that the impulse responses of the LDS approximate those of the predictor with exponentially small error.
Results
The main results include a theoretical guarantee on the prediction error of the distilled LDS, which decomposes into an exponentially small distillation term and an OSF learning term. Empirical evaluations on linear LDS benchmarks and MuJoCo behavior cloning demonstrate that the distilled LDS retains the behavior of the spectral predictor and often surpasses the performance of directly trained LDS baselines.
Implications
The findings suggest a new approach for modeling complex nonlinear systems in a more interpretable and computationally efficient manner. This has potential applications in various fields, including robotics, control systems, and any domain requiring accurate modeling of dynamical systems.
Matrix Zonotopic Attention: A Context-Adaptive Value Projection for Set Transformers
Theory
Efficient ML
Optimization
- Introduces Matrix Zonotopic Attention (MZAttn) for set transformers, enhancing adaptability to input sets.
- Defines Transformation Degrees of Freedom (TDOF) as a measure of complexity for target operators.
- Demonstrates that MZAttn can represent complex targets with a single layer, unlike standard attention which requires greater depth.
- Experimental results indicate significant performance improvements on high-complexity set-prediction tasks.
Read more
Matrix Zonotopic Attention: A Context-Adaptive Value Projection for Set Transformers
Summary
This paper introduces Matrix Zonotopic Attention (MZAttn), a novel attention mechanism designed for set transformers that addresses the limitations of standard multi-head attention in adapting to input sets. The authors analyze the asymmetry in standard attention, which combines input-dependent softmax routing with a fixed, input-independent value projection. They introduce the concept of Transformation Degrees of Freedom (TDOF) to quantify the complexity of target operators, demonstrating that context-rigid attention requires depth proportional to TDOF. In contrast, MZAttn employs a context-adaptive matrix-zonotope family for value projection, allowing it to represent complex targets with a single layer, thus reducing the required depth. The proposed architecture maintains permutation equivariance and provides a data-driven interpretation of reachability. Experimental results on various set-prediction tasks show that MZAttn significantly outperforms standard attention mechanisms, particularly on targets that exhibit high-rank, sparsely combinatorial dependencies on input sets, while showing competitive performance on simpler aggregate-statistic targets.
Methodology
The authors propose MZAttn, which replaces the fixed value projection in standard attention with a context-adaptive matrix-zonotope family. This involves a center matrix and generator matrices weighted by input-dependent gates. The architecture is evaluated through a series of experiments on set-prediction tasks, comparing its performance against standard attention mechanisms and other baselines.
Results
MZAttn outperformed standard multi-head attention on complex set-prediction tasks, achieving a significant RΒ² score of 0.69 on the minimum enclosing ball radius task, while standard attention methods remained below 0.025. The results validate the theoretical predictions regarding TDOF and highlight the selective advantage of MZAttn on tasks requiring high-rank, combinatorial input dependencies.
Implications
The findings suggest that MZAttn can be effectively applied in various domains requiring set-based processing, such as point-cloud analysis and molecular modeling. Its ability to adaptively project values based on input context could lead to advancements in tasks where traditional attention mechanisms struggle.
Why the Third Axis Is Freedom
Theory
Generative Models
Optimization
- Introduces 'freedom' as a measure of a model's generalization capability, surpassing generative expressivity.
- Proves that weaker models are more likely to generalize effectively, with empirical evidence supporting this claim.
- Demonstrates that increasing the number of outputs (K) in XM enhances model freedom and performance.
- Critiques generative expressivity for its limitations in ranking models and connecting to optimal generalization.
Read more
Why the Third Axis Is Freedom
Summary
This paper introduces the concept of 'freedom' as a critical factor in generative training models, particularly in the context of Explorative Modeling (XM). The author argues that while generative expressivity has been used to measure a model's capacity, it is fundamentally inferior to the measure of freedom, which reflects the weakness of constraints in a model's behavior. The paper presents a formal proof that weaker models are more likely to generalize effectively, with freedom selection outperforming minimum-description-length (MDL) by significant margins in induction experiments. The author empirically demonstrates that XM optimizes for freedom by showing that increasing the number of outputs (K) enhances the model's freedom and generalization capabilities. The paper also critiques the limitations of generative expressivity and establishes a connection between exploration in training and the concept of freedom, suggesting that XM serves as a practical method for achieving 'free-maxing' in model training. Overall, the findings indicate that selecting for freedom can lead to improved performance in generative models, especially under distribution shifts.
Methodology
The paper employs formal proofs and empirical experiments to establish the relationship between freedom and model generalization. It analyzes the performance of XM by varying the number of candidate outputs (K) and comparing the results against traditional measures like minimum-description-length (MDL). The author also derives mathematical identities related to best-of-K coverage and constructs a finite generative Stack Theory language to illustrate the concept of freedom.
Results
The author demonstrates that freedom selection outperforms MDL by 110-500% in induction experiments. Empirical results show that increasing K in XM leads to greater freedom and improved performance in 29 out of 30 validation cases. The paper establishes that the average XM loss is influenced by the probability of missing acceptable regions, with exploration enhancing match probabilities as K increases.
Implications
The findings suggest that focusing on freedom in model training can lead to more robust generative models, with potential applications in various domains such as computer vision, natural language processing, and reinforcement learning. The insights could influence future research directions in model architecture and training methodologies.
Perturbation Sensitivity at Convergence: A Simple Signal for Identifying Spuriously Correlated Samples
Computer Vision
Theory
Interpretability
- Introduces a method to identify spurious correlations in machine learning models post-convergence.
- Utilizes perturbation sensitivity to distinguish between SC and non-SC samples without group annotations.
- Demonstrates a significant improvement in worst-group accuracy on the Waterbirds dataset.
- Offers a simple, efficient procedure requiring only two forward passes per sample.
Read more
Perturbation Sensitivity at Convergence: A Simple Signal for Identifying Spuriously Correlated Samples
Summary
This paper addresses the challenge of identifying spurious correlations in machine learning models trained on empirical risk minimization. Such models often achieve high average accuracy but fail on subpopulations where the spurious correlation does not hold. Existing methods for identifying affected samples rely on early training signals, which require group annotations and can be difficult to implement. The author proposes a novel approach that utilizes perturbation sensitivity at convergence to identify samples that are spurious correlated (SC) versus non-spurious correlated (non-SC). The key insight is that SC samples are classified by a shared rule, while non-SC samples are memorized through individual configurations, making them more fragile. By applying a fixed perturbation to the inputs of a converged model, the author demonstrates that predictions for non-SC samples are more likely to change, thus allowing for effective identification of these samples. This method requires only two forward passes per training sample and does not necessitate group annotations or early stopping. The effectiveness of this approach is validated on the Waterbirds dataset, where rebalancing training using the identified samples significantly improves worst-group accuracy.
Methodology
The proposed method involves training a model to convergence and then applying a fixed perturbation (Gaussian blur) to each training sample. A binary sensitivity indicator is computed to determine if the perturbation changes the predicted class. This allows for the classification of samples into pseudo-groups based on their sensitivity to perturbation, facilitating the identification of non-SC samples for rebalancing training.
Results
The method successfully flagged 70.7% of minority samples and 3.5% of majority samples in the Waterbirds dataset, demonstrating a significant separation between SC and non-SC samples. Rebalancing training using the identified samples improved worst-group accuracy from 57.3% to 80.8%, compared to 85.8% accuracy using ground-truth group labels.
Implications
This approach provides a practical solution for identifying and mitigating the effects of spurious correlations in machine learning models, particularly in scenarios where group annotations are unavailable. It enhances model reliability across subpopulations, making it applicable in various real-world settings where subgroup performance is critical.
A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance
Theory
Efficient ML
Multimodal
- Introduces a six-dimensional taxonomy for post-training adaptation techniques in machine learning.
- Clarifies terminological ambiguities and relationships among various adaptation methods.
- Maps 48 adaptation techniques to the proposed taxonomy, aiding in technical documentation and governance.
- Identifies open challenges in evaluation, reproducibility, and governance-aware adaptation workflows.
Read more
A Six-Dimensional Taxonomy of Post-Training Adaptation Techniques with Applications in AI Governance
Summary
This paper addresses the fragmented landscape of post-training adaptation techniques in machine learning, which are crucial for modifying models after deployment. The authors propose a six-dimensional taxonomy that organizes these techniques based on six criteria: mechanism, goal, data requirements, persistence, structural scope, and model type. This taxonomy helps clarify terminological ambiguities and cross-dimensional relationships that currently hinder effective comparison and governance of adaptation methods. The paper identifies 48 distinct adaptation techniques and maps them to the six dimensions, providing a comprehensive framework that can support technical documentation and compliance with regulatory standards. Additionally, the authors highlight the challenges in evaluating and reproducing adaptation methods, as well as the need for governance-aware workflows in AI development. The taxonomy serves as a foundational tool for understanding and managing the complexities of post-training adaptations in the context of evolving regulatory landscapes.
Methodology
The authors conducted a comprehensive survey of existing literature on post-training adaptation techniques, identifying and categorizing 48 techniques into a six-dimensional framework. They analyzed the relationships among these techniques and their implications for AI governance, providing a detailed mapping of each technique to the taxonomy.
Results
The proposed taxonomy effectively organizes the diverse landscape of post-training adaptation techniques, resolving issues of terminological ambiguity and model-type conflation. The centerpiece table in the paper maps all identified techniques to their respective six-dimensional profiles, facilitating better understanding and governance of model adaptations.
Implications
The taxonomy can enhance technical documentation, model-change tracking, and compliance with regulatory requirements in AI governance. It provides a structured approach to understanding how adaptations can impact model behavior and regulatory classification, which is crucial for developers and policymakers in the AI field.
Accelerating nanodrug development in continuous flow systems using informed prediction models based on low-cost surrogate nanoparticles
Optimization
- Introduction of a predictive modeling approach for nanoparticle development.
- Validation of shape-constrained models using minimal empirical data.
- Demonstrated reduction in experimental workflows for nanoparticle characterization.
- Emphasis on the transition from empirical to data-driven design in nanomedicine.
Read more
Accelerating nanodrug development in continuous flow systems using informed prediction models based on low-cost surrogate nanoparticles
Summary
This paper addresses the challenges in the development of nanotherapeutics, which often require extensive empirical optimization due to the sensitivity of nanoparticle properties to minor changes in process parameters. The authors propose a predictive modeling approach that incorporates shape constraints to enhance the estimation of nanoparticle characteristics across various production conditions. By utilizing controlled microfluidic methods, they systematically prepared liposomes and lipid nanoparticles under varying conditions and validated their shape-constrained model with minimal empirical data. The results indicate that this modeling framework significantly reduces the need for extensive experimental workflows, facilitating more efficient and rational development processes for nanomedicine systems. The study highlights the transition from empirical discovery to data-driven design in the field of nanocarrier-based delivery systems, emphasizing the importance of predictive models in optimizing critical quality attributes such as particle size and polydispersity index (PDI).
Methodology
The authors employed controlled microfluidic methods to prepare liposomes and lipid nanoparticles under varying lipid concentrations, flow rates, and aqueous-to-organic mixing ratios. They developed a shape-constrained predictive model informed by experimental data and expert knowledge, which was validated for pharmaceutical applications.
Results
The shape-constrained modeling approach successfully predicted nanoparticle size and dispersity, demonstrating a significant reduction in the need for extensive empirical testing. This predictive capability supports more efficient process development in the manufacturing of nanomedicine systems.
Implications
The findings suggest that integrating predictive modeling into the development of nanotherapeutics can streamline the optimization process, reduce costs, and accelerate the time to market for new nanodrug formulations. This approach could lead to more reliable and scalable production methods in the pharmaceutical industry.
Timestep-Conditioned Transformers for Global Weather Forecasting
Time Series
- GEM-3 allows for configurable multi-timestep inference, enabling dynamic adjustment of forecasting timesteps at inference time.
- The model employs anomaly-space modeling to improve stability and reduce drift in long-range forecasts.
- GEM-3 outperforms traditional models and achieves near-state-of-the-art CRPS scores across multiple variables.
- The architecture includes advanced features such as neighborhood attention and efficient training strategies.
Read more
Timestep-Conditioned Transformers for Global Weather Forecasting
Summary
This paper presents GEM-3, a novel probabilistic global weather forecasting model that addresses the limitations of existing autoregressive models by allowing configurable multi-timestep inference. Traditional weather forecasting models typically operate on fixed timesteps, which present a trade-off between short-term accuracy and long-term stability. GEM-3 enables users to dynamically adjust the timestep during inference, thus balancing the need for detailed short-range forecasts with the stability required for extended-range predictions. The model employs a lightweight neighborhood-attention transformer architecture with approximately 134 million parameters and incorporates advanced techniques such as anomaly-space modeling to enhance stability over long forecasting horizons. The authors demonstrate that GEM-3 achieves near-state-of-the-art performance in continuous ranked probability scores (CRPS) and outperforms existing models like ECMWF's ENS across various variables and lead times, while also providing reliable extended-range forecasts up to 126 days.
Methodology
The authors developed GEM-3 using a transformer architecture that incorporates configurable multi-timestep inference and anomaly-space modeling. The model operates on an equirectangular grid and utilizes neighborhood attention blocks to enhance locality bias. It is trained on a diverse dataset, allowing it to adaptively switch between timesteps during inference, thus optimizing for both short-term and long-term forecasting needs.
Results
GEM-3 achieves competitive CRPS scores, outperforming ECMWF's ENS across all analyzed variables and lead times. It shows comparable performance to other leading ML models while providing significant advantages in sub-timestep diagnostics. The model also maintains stability and skill in extended-range rollouts, outperforming ECMWF's extended-range ensemble forecasts up to 46 days and converging towards climatological baseline scores out to 126 days.
Implications
The development of GEM-3 has significant implications for operational weather forecasting, allowing for more flexible and accurate predictions that can adapt to varying forecasting needs. This model could enhance decision-making in sectors reliant on accurate weather forecasts, such as agriculture, disaster management, and climate research.
An Optimal Agnostic PAC Algorithm
Theory
- The paper introduces a deterministic learner that achieves optimal risk bounds in agnostic PAC learning.
- The learner's performance is characterized by a high probability bound that matches established lower bounds.
- A new class-dependent edge isoperimetric inequality is a key component of the methodology.
- The results provide a comprehensive understanding of sample complexity in agnostic PAC learning.
Read more
An Optimal Agnostic PAC Algorithm
Summary
This paper presents a novel learner that achieves the statistically optimal risk bound in the context of agnostic PAC (Probably Approximately Correct) learning. The authors construct a deterministic learner that operates without prior knowledge of the optimal risk L* or the confidence parameter Ξ΄, ensuring that for any distribution and sample size n, the learner's output classifier bh satisfies a high probability bound on the excess risk. The proposed algorithm's performance is characterized by a bound that matches the lower bounds established by previous works, thus settling the sample complexity of agnostic PAC learning up to universal constants. The methodology leverages a new class-dependent edge isoperimetric inequality and suffix averaging techniques to derive the optimal leave-one-out bound, which is crucial for the learner's performance. The results indicate that the proposed learner not only meets but also optimally aligns with the theoretical limits of agnostic PAC learning, providing a significant advancement in the understanding of sample complexity in this domain.
Methodology
The authors develop a deterministic learner based on a new edge isoperimetric inequality that controls induced edge counts through approximation errors. The methodology includes suffix averaging with variance control, adapted from previous works, to derive the optimal leave-one-out bound, which is then transformed into a PAC learner.
Results
The main result is a high probability bound on the excess risk of the proposed learner, which is shown to be optimal and matches the lower bounds established in prior literature. Specifically, the bound is expressed in terms of the optimal risk L*, VC dimension d, and sample size n, confirming the learner's efficacy across various distributions.
Implications
The findings have significant implications for the field of statistical learning theory, particularly in improving the understanding of sample complexity in agnostic settings. The optimal learner can be applied to various classification tasks, enhancing the performance of machine learning models in real-world applications where prior knowledge of the optimal risk is unavailable.
Hybrid Probabilistic Zonotopes for Identifiable and Refinable Predictive Uncertainty
Time Series
Theory
Robotics
- HProbZ effectively separates and identifies three sources of uncertainty in predictions.
- The method allows for observation-driven refinement of predictive distributions in a single forward pass.
- HProbZ densities are representationally distinct from finite Gaussian mixtures.
- Empirical results demonstrate superior performance in trajectory prediction tasks compared to existing methods.
Read more
Hybrid Probabilistic Zonotopes for Identifiable and Refinable Predictive Uncertainty
Summary
The paper introduces the Hybrid Probabilistic Zonotope (HProbZ), a novel output head for neural networks that effectively captures three distinct sources of uncertainty in predictive tasks: discrete modal choice, bounded systematic drift, and irreducible stochastic noise. Traditional probabilistic prediction heads, such as Gaussian mixtures or single conformal regions, fail to adequately separate these uncertainties. HProbZ employs a combination of binary, bounded, and stochastic generators to represent these uncertainties, allowing for a closed-form likelihood through convolution. A key feature of HProbZ is the sharing of the bounded generator across prediction steps, which enables observation-driven refinement of the predictive distribution in a single forward pass. The authors demonstrate that the generators are identifiable from the likelihood and that HProbZ densities are distinct from finite Gaussian mixtures. The empirical analysis on trajectory prediction benchmarks shows that HProbZ outperforms existing methods, providing structural properties that are not jointly available in other approaches.
Methodology
The authors define the Hybrid Probabilistic Zonotope (HProbZ) as a structured output head that combines binary, bounded, and stochastic generators. These generators are trained jointly under a single closed-form convolution likelihood. The bounded generator is shared across prediction steps, allowing for simultaneous updates to the predictive distribution as observations are made.
Results
Empirical evaluations on trajectory forecasting benchmarks indicate that HProbZ achieves a minimum average displacement error (minADE) of 0.20 at 20 steps, outperforming same-encoder mixture baselines such as MDN-K20, which achieved a minADE of 0.32. HProbZ also demonstrated a 30% reduction in trajectory-level collisions compared to MDN-K2, highlighting its effectiveness in refining predictive uncertainty.
Implications
The HProbZ framework can significantly enhance predictive modeling in applications requiring accurate uncertainty quantification, such as autonomous driving and robotics. Its ability to refine predictions based on observations in real-time could lead to more reliable decision-making systems in dynamic environments.
PRISM: Priority-aware Rubric Internalization via Structured Multimodal Data Synthesis
Multimodal
Large Language Models
NLP
- PRISM redefines rubric comprehension as an executor-side task, enhancing multimodal instruction following.
- The framework synthesizes structured, priority-aware training data to improve model performance.
- PRISM-Eval provides a novel evaluation method that does not rely on external judge models.
- Significant accuracy improvements were observed with minimal synthesized data, demonstrating scalability.
Read more
PRISM: Priority-aware Rubric Internalization via Structured Multimodal Data Synthesis
Summary
The paper introduces PRISM, a novel framework designed to enhance multimodal instruction comprehension by addressing the supervision gap in existing datasets. Traditional multimodal models often treat instructions as single questions, failing to account for the varying importance of multiple requirements. PRISM reframes the task as 'rubric comprehension,' where the model acts as an executor that verifies each rule in a prioritized rubric before producing an overall judgment. The authors propose a four-stage data synthesis framework that generates persona-task pairs, prefix-guided rule sets, quality-filtered rubrics, and structured verification traces. Additionally, they introduce PRISM-Eval, an evaluation protocol that allows for automatic assessment of model performance without requiring external judge models. The results demonstrate that with only 10,000 synthesized samples, PRISM significantly improves the accuracy of the Qwen3-VL-4B model on PRISM-Eval, achieving a Strict accuracy increase from 9.5% to 30.1%, while maintaining performance on general benchmarks. This improvement is also transferable to four other open-source multimodal large language models (MLLMs), indicating the scalability of structured rubric supervision for multi-rule, priority-aware instruction following.
Methodology
The methodology involves a four-stage data synthesis framework that includes persona-task discovery, prefix-guided rule completion, rule quality assessment, and structured response generation. This structured approach allows the model to internalize multi-rule instructions effectively. The evaluation is conducted using PRISM-Eval, which employs Loose and Strict metrics for assessing model performance without the need for external judges.
Results
The implementation of PRISM resulted in a substantial increase in Strict accuracy for the Qwen3-VL-4B model, rising from 9.5% to 30.1% on PRISM-Eval. This improvement was achieved with only 10,000 synthesized samples and was consistent across four additional MLLMs, showcasing the effectiveness of the structured rubric supervision.
Implications
The findings suggest that structured rubric supervision can significantly enhance the capability of multimodal models to follow complex, priority-aware instructions. This has potential applications in various domains, including product quality inspection and advertisement compliance review, where nuanced understanding of instructions is critical.