AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
48
Papers today
8h
Update frequency
7
Days of history
Learning to Run Power Networks: Effective AlphaZero-inspired Topological Control
Reinforcement Learning
- AlphaZero-inspired methods can significantly enhance the stability of power networks under strain from renewable energy sources.
- An optimized approach achieves a peak survivability of 98.43%, outperforming traditional RL methods like PPO.
- MCTS without prior policy guidance can improve training efficiency.
- Simple binary rewards are more effective for guiding search than complex reward structures.
Read more
Learning to Run Power Networks: Effective AlphaZero-inspired Topological Control
Summary
This paper addresses the challenges posed by the integration of volatile renewable energy sources into power grids, which can lead to instability and congestion. It explores the use of Reinforcement Learning (RL) for autonomous topological reconfiguration of power networks, presenting a model-based approach inspired by AlphaZero that employs Monte Carlo Tree Search (MCTS) for proactive grid management. The authors systematically evaluate the impact of reward functions, observation density, and search guidance on the agent's performance. The study finds that an optimized AlphaZero approach achieves a peak survivability of 98.43%, significantly surpassing the performance of a Proximal Policy Optimization (PPO) variant. Notably, the research highlights that MCTS can be more efficient when conducted without prior policy guidance and that simple binary rewards can provide better search guidance than complex multi-objective functions. The findings suggest that while AlphaZero is effective for topological control, a successful implementation requires a minimalist integration of domain-specific heuristics, binary rewards, and a restricted observation space.
Methodology
The authors utilize a model-based AlphaZero-inspired framework that combines deep neural networks with Monte Carlo Tree Search (MCTS) to explore the combinatorial action space of power grid configurations. They conduct systematic evaluations of various design factors, including reward structures, observation density, and search guidance, using the IEEE-14 bus system as a controlled environment.
Results
The optimized AlphaZero approach achieved a peak survivability of 98.43%, significantly outperforming the PPO variant. The study also found that conducting MCTS without prior policy guidance enhances training efficiency and that binary survival rewards provide more effective search guidance than complex reward functions.
Implications
The findings suggest that integrating AlphaZero-inspired methods into power grid management can lead to more efficient and stable operations, particularly as renewable energy sources become more prevalent. This approach could help mitigate the risks of grid congestion and cascading failures, contributing to the reliability of energy systems in the transition to carbon neutrality.
Finding the Needle in a Haystack: Test-Time Analog Circuit Representation Adaptation for Bayesian Optimization
Optimization
- TTARO introduces an online adaptation mechanism for circuit representations during Bayesian optimization.
- The framework allows for continuous learning and adjustment of circuit embeddings based on newly observed performance metrics.
- TTARO outperforms traditional fixed representation methods and DKL in terms of optimization efficiency.
- The approach is compatible with various acquisition functions and Gaussian-process kernels.
Read more
Finding the Needle in a Haystack: Test-Time Analog Circuit Representation Adaptation for Bayesian Optimization
Summary
This paper presents Test-Time Analog Representation Adaptation for Bayesian Optimization (TTARO), a novel online deep-kernel Bayesian optimization framework aimed at improving the efficiency of analog circuit topology search. Traditional Bayesian optimization methods often rely on fixed circuit embeddings, which can misalign with the optimization objectives, leading to suboptimal search performance. TTARO addresses this issue by adapting circuit representations dynamically as new figure-of-merit (FoM) observations are collected during the optimization process. By jointly learning a nonlinear feature transformation and a Gaussian-process surrogate model, TTARO continuously refines the representation of circuit topologies, ensuring that the search space aligns more closely with the optimization goals. The authors conducted extensive experiments across various circuit encoders, kernels, and acquisition functions, demonstrating that TTARO significantly outperforms conventional methods, including fixed embedding approaches and Deep Kernel Learning (DKL). The results indicate that TTARO reduces regret area under the curve (AUC) by an average of 15.2% compared to standard Bayesian optimization and by 20.7% relative to DKL, showcasing its effectiveness in optimizing analog circuit designs.
Methodology
TTARO employs an online deep-kernel Bayesian optimization framework that adapts pretrained circuit embeddings by learning a nonlinear feature transformation and a Gaussian-process surrogate model. This adaptation occurs iteratively as new FoM observations are gathered, allowing the optimization process to align more closely with the objectives being pursued.
Results
TTARO demonstrated a reduction in regret AUC by 15.2% on average compared to conventional Bayesian optimization and 20.7% compared to DKL across 40 different configurations. In some settings, TTARO achieved reductions in regret as large as 46.7%, indicating substantial improvements in optimization performance.
Implications
The findings suggest that dynamic adaptation of representations can significantly enhance the efficiency of optimization processes in analog circuit design, potentially leading to faster and more effective design cycles in electronic design automation. This approach could be extended to other areas requiring optimization under similar constraints.
From BERT to Frontier Agents: Eight Years of Language-Model Progress, the Collapse of the Capability-Cost Curve, and the Rise of Task-Targeted Models
NLP
Large Language Models
Efficient ML
- Significant growth in language model capabilities from BERT to trillion-parameter systems.
- Collapse of the capability-cost curve, with budget models outperforming flagship models at lower costs.
- Emergence of task-fragmented models, necessitating specialized routing for optimal performance.
- Introduction of a confidence estimation model to improve prediction accuracy in task-specific scenarios.
Read more
From BERT to Frontier Agents: Eight Years of Language-Model Progress, the Collapse of the Capability-Cost Curve, and the Rise of Task-Targeted Models
Summary
This paper provides an empirical review of the evolution of language models from BERT in 2018 to advanced agentic systems by 2026. It highlights significant advancements in model capabilities, cost efficiency, and task specialization over eight years. The author traces the progression through four paradigms: fine-tuning, in-context learning, alignment, and reasoning. Key findings include a 5.8x annual growth in the odds of solving tasks on public benchmarks, a dramatic decline in input-token costs (approximately 60x), and the emergence of task-targeted models that outperform previous generations at lower costs. The study also introduces a two-model routing approach that optimizes performance across tasks and presents a confidence estimation model for enhancing prediction accuracy. The paper concludes with a comprehensive repository of data and methods for reproducibility.
Methodology
The author conducted a longitudinal analysis of language models from 2018 to 2026, utilizing public benchmarks and pricing data to measure performance and cost efficiency. The study involved empirical experiments, including a grade-school-math evaluation and the development of a confidence model for prediction ranking. A two-model routing strategy was also implemented to assess performance gains across tasks.
Results
The analysis revealed a 5.8x annual increase in the likelihood of solving tasks on SWE-bench Verified benchmarks. The cost analysis indicated a 60x reduction in input-token prices from GPT-3 to GPT-5.6 Luna, with budget models achieving comparable performance to flagship models. The task-level results showed a fragmented landscape, with different models excelling in specific areas. The confidence model demonstrated effective triage capabilities, achieving high accuracy in its top predictions.
Implications
The findings suggest that the rapid advancement of language models can lead to more efficient and specialized applications in various domains, including software development and mathematical problem-solving. The introduction of task-targeted models and routing strategies could enhance the deployment of AI systems in real-world scenarios, making them more accessible and cost-effective.
CAKE: Compiler-Agent Co-Design for Frontier Kernel Evolution
Optimization
Efficient ML
Theory
- CAKE integrates compiler and agent co-design to improve GPU kernel evolution.
- The framework uses a typed intermediate representation (IR) for better hardware-specific control.
- Localized correctness and performance diagnostics are provided to agents, enhancing decision-making.
- Significant performance improvements were observed in benchmarks compared to traditional methods.
Read more
CAKE: Compiler-Agent Co-Design for Frontier Kernel Evolution
Summary
The paper introduces CAKE, a novel framework that integrates compiler and agent co-design to enhance the evolution of GPU kernels. Traditional approaches treat the compiler as a black box, limiting the feedback available to agents when optimizing kernels. CAKE addresses this by allowing agents to work with a typed intermediate representation (IR) that provides detailed insights into hardware-specific scheduling and performance diagnostics. This enables agents to make informed decisions during the kernel evolution process. The framework evolves both the kernels and the compiler, allowing for continuous improvement based on performance feedback and correctness diagnostics. The authors demonstrate CAKE's effectiveness through various benchmarks, showing significant performance improvements over traditional CUDA/PTX implementations. The results indicate that CAKE can facilitate the development of high-performance GPU kernels that are tailored to specific workloads, thereby bridging the gap between expert kernel programming and automated kernel generation.
Methodology
The authors developed CAKE by creating a typed intermediate representation (IR) that allows agents to edit and optimize GPU kernels with explicit hardware scheduling. The framework includes a feedback loop where the compiler provides detailed diagnostics on correctness and performance, enabling agents to refine their approaches iteratively. The methodology was validated through a series of benchmarks across various kernel families, demonstrating the effectiveness of the CAKE framework in real-world scenarios.
Results
In benchmarks, the best candidate using CAKE IR achieved a median performance of 1.144 times the tuned FlashML baseline, compared to 0.928 times for direct CUDA/PTX. Additionally, the agent-generated Kimi Delta Attention achieved a geometric mean speedup of 2.05 times over the official FlashKDA. Performance improvements for dispatcher-backed KNN and KMeans families ranged from 1.42 to 2.12 times across over 400 shapes, with four kernel changes made available as upstream pull requests.
Implications
The CAKE framework has the potential to revolutionize GPU programming by enabling more efficient kernel development and optimization. It can lead to the creation of high-performance kernels that are better suited for specific applications, ultimately enhancing the capabilities of GPU computing in various domains such as machine learning, scientific computing, and real-time data processing.
Prof-K: Probabilistic One-Pass Filtering for Efficient Top-k Selection
Efficient ML
Optimization
Theory
- Prof-K provides a fast, scalable, and distribution-agnostic solution for top-k selection.
- The algorithm guarantees high-probability correctness and buffer size through a probabilistic filtering approach.
- Empirical results show 1.5Γβ10Γ speedups over existing top-k implementations, particularly beneficial for large datasets.
- Prof-K allows for flexible accuracy-speed trade-offs, enabling users to tune performance based on their requirements.
Read more
Prof-K: Probabilistic One-Pass Filtering for Efficient Top-k Selection
Summary
The paper introduces Prof-K, a novel algorithm designed for efficient top-k selection, which is crucial in various computational fields such as databases and machine learning. Traditional methods for top-k selection face inefficiencies as data sizes increase, with exact methods being computationally expensive and approximate methods often failing under adversarial conditions. Prof-K addresses these challenges by employing a probabilistic one-pass filtering approach. It begins with a small random sample to estimate an adaptive threshold, allowing the algorithm to stream through the input data only once and retain elements above this threshold in a compact buffer. An exact top-k selection is then performed on this buffer, ensuring that the true top-k elements are recovered with a high probability, independent of the input distribution. The authors provide theoretical guarantees for the algorithm's correctness and buffer size, along with an optimal sample size that minimizes overhead. Empirical evaluations demonstrate that Prof-K achieves significant speedups (1.5Γβ10Γ) over existing optimized implementations, particularly in scenarios with large data sizes and small to moderate k values. The algorithm also allows for flexible accuracy-speed trade-offs, enabling users to adjust parameters based on their specific needs. Additionally, Prof-K shows promise in reducing training times for BatchTopK Sparse Autoencoders without compromising performance.
Methodology
Prof-K utilizes a one-pass filtering mechanism where a small random sample is used to estimate a threshold. The algorithm streams through the input data once, retaining only those elements above the threshold in a compact buffer, followed by an exact top-k selection from this buffer. The guarantees of the method are derived from the ranks induced by uniform sampling, making it robust against various input distributions.
Results
Prof-K achieves speed improvements of 1.5Γ to 10Γ compared to highly optimized implementations like PyTorch topk and RadiK, especially in large-scale scenarios with small to moderate k values. The algorithm maintains high accuracy and allows for a flexible trade-off between speed and exact recovery.
Implications
The implications of Prof-K are significant for applications requiring efficient top-k selection, particularly in large-scale machine learning workloads and real-time data processing. Its robustness to input distribution makes it suitable for adversarial settings, enhancing the reliability of systems that depend on top-k operations.
QUASAR: Lowering the Loss Floor of Quantization-Aware Training with Loss-Aware Reconstruction
NLP
Large Language Models
Optimization
- QUASAR continuously minimizes loss-aware reconstruction error during QAT to improve training trajectories.
- The method uses exponential moving averages for saliency estimation and optimizes clipping ranges and dequantizers in real-time.
- Empirical results show QUASAR outperforms traditional QAT methods, achieving lower KL divergence and higher accuracy across multiple tasks.
Read more
QUASAR: Lowering the Loss Floor of Quantization-Aware Training with Loss-Aware Reconstruction
Summary
The paper introduces QUASAR, a novel method for quantization-aware training (QAT) aimed at improving model quality during low-precision inference. Traditional QAT methods face challenges due to a structural mismatch in optimization, where the loss is computed using lossy reconstructions of full-precision weights while updates are applied to the latent weights. This leads to a higher loss floor. QUASAR addresses this by continuously performing lightweight, loss-aware reconstruction during training, thus optimizing the reconstruction process iteratively. The method employs exponential moving averages of squared gradients to estimate saliency, searches over clipping ranges, and fits affine dequantizers using saliency-weighted least squares. The theoretical analysis shows that minimizing the loss-aware reconstruction error directly influences the QAT convergence bound and the final quantized model's loss. Empirical results demonstrate that QUASAR significantly reduces KL divergence and improves accuracy across various tasks compared to existing QAT methods, achieving notable performance gains at low bit widths.
Methodology
QUASAR integrates loss-aware reconstruction into the QAT process by decomposing reconstruction into quantization and dequantization stages. It uses an exponential moving average of squared gradients to estimate saliency, searches over candidate clipping ranges, and applies saliency-weighted least squares to fit dequantizers, optimizing the reconstruction error iteratively during training.
Results
QUASAR achieved the lowest held-out KL divergence among competitive QAT methods at 2, 3, and 4 bits, with reductions of at least 10% at 3 and 4 bits and 29% at 2 bits. It improved average accuracy by 3.5β4.3 percentage points over strong QAT and PTQ baselines at 2 bits and reduced held-out KL by approximately 30% when applied to NVFP4.
Implications
The findings suggest that QUASAR can enhance the deployment of low-precision models in large language applications, improving efficiency without sacrificing quality. This has significant implications for the cost-effectiveness of deploying large language models in real-world applications.
Robust data-driven discovery of fractional differential equations via weak formulations and Pareto-based subset selection
Theory
- Weak-Pareto combines weak formulations with Pareto-based selection for robust FPDE discovery.
- The framework effectively mitigates noise amplification issues associated with fractional differentiation.
- Weak-Pareto outperforms traditional methods in recovering equations from noisy data.
- The methodology allows for continuous-order optimization, avoiding discretization biases.
Read more
Robust data-driven discovery of fractional differential equations via weak formulations and Pareto-based subset selection
Summary
This paper presents Weak-Pareto, a novel framework for discovering fractional partial differential equations (FPDEs) from noisy data. The challenge in this domain arises from the amplification of high-frequency noise due to fractional differentiation and the unknown nature of derivative orders. Weak-Pareto integrates an adjoint-consistent weak formulation of fractional terms with a Pareto-based subset selection approach, allowing for effective handling of both linear and nonlinear terms. The methodology replaces traditional pointwise differentiation with smoothing integration, enhancing noise robustness. Coefficients are fitted using ridge regression within a differential-evolution search that accounts for the continuous nature of derivative orders. The framework demonstrates superior performance in recovering parsimonious structures from both clean and noisy measurements across various benchmarks, including fractional advection-diffusion and Burgers equations. The results indicate that Weak-Pareto retains the correct support even under high noise levels, outperforming traditional strong-form methods. Additionally, a two-dimensional extension of Weak-Pareto successfully identifies coordinate-dependent orders, showcasing its versatility in complex scenarios.
Methodology
The Weak-Pareto framework employs an adjoint-consistent weak formulation for fractional operators, replacing pointwise differentiation with integral measurements. It utilizes ridge regression for coefficient fitting and a differential-evolution search for continuous-order optimization. The Pareto-based subset selection enhances model robustness by selecting features based on validation error and complexity.
Results
Weak-Pareto successfully recovers parsimonious structures from both clean and noisy data across various benchmarks. It maintains correct support at all tested noise levels, significantly outperforming strong-form counterparts. The framework also shows improved operator recovery consistency and reduced runtime compared to neural network baselines.
Implications
The proposed framework has significant implications for data-driven modeling of complex systems described by fractional differential equations, which are prevalent in fields such as physics, biology, and finance. Its robustness against noise makes it a valuable tool for real-world applications where data quality is often compromised.
RecipeNet: A Hierarchical Transformer for Recipe Data
NLP
- RecipeNet effectively models the hierarchical structure and sequential dependencies of recipe data.
- The architecture outperforms traditional tabular learning methods by preserving the native recipe structure.
- RecipeNet consists of three stages: field-level tokenization, step-level fusion, and recipe-level sequence encoding.
- Experimental results highlight the importance of hierarchical modeling for effective recipe representation learning.
Read more
RecipeNet: A Hierarchical Transformer for Recipe Data
Summary
RecipeNet introduces a novel hierarchical Transformer architecture designed specifically for recipe data, which is prevalent in various domains such as materials synthesis and pharmaceutical formulation. Traditional tabular learning methods struggle with recipe data due to its unique characteristics, including variable schemas, hierarchical structures, and sequential dependencies. RecipeNet addresses these challenges by preserving the inherent structure of recipes, allowing for the encoding of field-level interactions within each procedural step and capturing sequential dependencies across steps. The architecture consists of three main stages: field-level tokenization, step-level fusion, and recipe-level sequence encoding. Experimental evaluations across multiple recipe datasets demonstrate that RecipeNet consistently outperforms existing tabular models, showcasing the effectiveness of hierarchical and sequential modeling in recipe representation learning.
Methodology
RecipeNet employs a hierarchical Transformer architecture that includes field-level tokenization to process heterogeneous fields within each procedural step, followed by step-level fusion to integrate information from these fields, and finally recipe-level sequence encoding to capture dependencies across the procedural steps. This approach allows RecipeNet to maintain the original structure of recipes without flattening them into fixed-schema representations.
Results
The experimental results indicate that RecipeNet consistently outperforms existing tabular models across multiple recipe datasets and tasks, demonstrating its superior capability in capturing both local field interactions and global procedural context.
Implications
The development of RecipeNet has significant implications for various industries that rely on recipe data, such as manufacturing, pharmaceuticals, and cooking. By improving the representation learning of recipes, RecipeNet can enhance predictive modeling, optimize processes, and facilitate better decision-making in these domains.
ProME: Prototype-Margin Environments with Repair-Aware Selection for Group-Robust Learning
Theory
Optimization
- ProME aligns environment construction and model selection with the deployed predictor without requiring training-group labels.
- The framework introduces a median split of prototype margins to create balanced environments for invariant learning.
- Theoretical bounds on worst risks are established, showing transferability to oracle groups under alignment conditions.
- ProME demonstrates superior performance in worst-group accuracy compared to existing methods across multiple benchmarks.
Read more
ProME: Prototype-Margin Environments with Repair-Aware Selection for Group-Robust Learning
Summary
This paper addresses the challenge of group-robust learning, particularly in scenarios where training-group labels are unavailable. The authors propose a novel framework called ProME (Prototype-Margin Environments) that formulates the problem as endogenous environments with repair-aware selection (ERAS). ProME aims to align environment construction and model selection with the deployed predictor, thereby enhancing the accuracy of rare subpopulations. The methodology involves splitting prototype margins at their median to create balanced environments and fitting a group-balanced linear head on validation data to rank predictors based on worst-group accuracy (WGA). Theoretical bounds on the worst risk across inferred environments are established, demonstrating that these bounds can transfer to oracle groups under specific conditions. Extensive experiments on benchmarks such as Waterbirds, CelebA, and CivilComments reveal that ProME outperforms existing methods in achieving higher average worst-group accuracy while effectively addressing shortcut-conflicting examples and model selection misalignments.
Methodology
ProME constructs environments by assigning prototype margins between observed classes and competing classes using a cosine-prototype classifier. It employs a median split of these margins to create approximately balanced environments. In the second stage, it ranks encoder checkpoints based on validation worst-group accuracy after classifier repair, ensuring alignment with the deployed predictor.
Results
ProME achieves the highest average worst-group accuracy among compared methods on benchmarks like Waterbirds, CelebA, and CivilComments. The experiments demonstrate that the proposed method effectively enriches shortcut-conflicting examples and reshapes candidate evaluations, leading to improved group robustness.
Implications
The findings suggest that ProME can be applied in various domains where group robustness is critical, such as medical diagnosis and content moderation, particularly in situations with imbalanced group representations. The methodology can enhance the reliability of machine learning models in real-world applications by ensuring equitable performance across diverse subpopulations.
Fast A/B/n Testing: Exact Multi-Policy Comparison via Tree-Coupled Feedback Sharing
Theory
Efficient ML
Optimization
- Introduces Tree-Coupled A/B Testing (TCAB) for efficient multi-policy comparisons.
- Reduces the number of reward-bearing interactions by sharing outcomes among policies.
- Maintains each policy's standalone trajectory distribution despite shared feedback.
- Achieves conditionally optimal query costs compared to traditional independent A/B/n testing.
Read more
Fast A/B/n Testing: Exact Multi-Policy Comparison via Tree-Coupled Feedback Sharing
Summary
This paper introduces Tree-Coupled A/B Testing (TCAB), a novel method for conducting A/B/n tests that allows for exact multi-policy comparisons while sharing feedback efficiently. Traditional A/B/n testing requires independent trajectories for each policy, leading to a high number of reward-bearing interactions. TCAB addresses this by coupling the trajectories of policies through a predictable tree structure, enabling shared outcomes for policies that would make the same decision in identical contexts. The method ensures that each policy retains its standalone trajectory distribution while minimizing redundant interactions, thus reducing the total number of reward queries. The paper establishes that the expected query cost for TCAB is T plus the cumulative total variation over the selected tree edges, which is significantly lower than the JT cost of independent runs. Experimental results demonstrate that TCAB achieves substantial improvements in the cost-precision frontier across various applications, including reward-model evaluation and adaptive search policies.
Methodology
The TCAB method utilizes a tree structure to couple the trajectories of multiple policies. At each round, it samples from a tree that connects the histories of the policies, allowing for shared outcomes when policies make the same decision. The algorithm operates in a round-synchronous manner, enabling efficient parallelism in querying outcomes and updating policies.
Results
The expected number of reward queries in TCAB is T plus the cumulative total variation over the tree edges, leading to a cost of E[N(T)] = T + o(T) for fixed J, which is significantly lower than the JT cost associated with independent A/B/n testing. The experiments conducted show substantial improvements in cost-effectiveness and precision in various applications.
Implications
The TCAB framework can be applied in online experimentation across various digital platforms, particularly in scenarios where multiple adaptive decision policies are evaluated. It offers a more efficient way to conduct A/B/n tests, potentially leading to faster iterations and better resource utilization in machine learning evaluations.
Adaptive Protection for Evolutionary Feature Construction in Symbolic Regression with Application to Credit Classification
Optimization
Theory
- Introduction of an adaptive protection mechanism for preserving important features during evolutionary feature construction.
- Utilization of feature importance metrics to guide the preservation process.
- Empirical evaluation demonstrates improved solution quality across multiple datasets and base learners.
- The approach is applicable to both symbolic regression and credit classification tasks.
Read more
Adaptive Protection for Evolutionary Feature Construction in Symbolic Regression with Application to Credit Classification
Summary
This paper addresses the challenge of preserving important constructed features during evolutionary feature construction in symbolic regression. Traditional methods often fail to maintain valuable features when genetic operators modify them, leading to a degradation in solution quality. The authors propose an adaptive protection mechanism that utilizes feature importance metrics to selectively preserve constructed features based on their significance. This mechanism allows for stronger protection of more important features while permitting less important ones to be modified, thus facilitating the generation of new feature combinations. The authors evaluate their approach using various feature importance calculation methods across different base learners and demonstrate its effectiveness through experiments on 98 regression benchmark datasets and two credit classification datasets. The results indicate that the proposed mechanism consistently enhances solution quality compared to baseline approaches, showcasing its robustness and applicability beyond symbolic regression.
Methodology
The authors developed an adaptive protection mechanism that integrates feature importance metrics to guide the preservation of constructed features during evolutionary processes. This mechanism is embedded in a universal protection framework that can be applied to any genetic operator, ensuring that more important features receive stronger protection while allowing less important features to be modified or replaced.
Results
Experimental results on 98 regression benchmark datasets showed that the proposed adaptive protection mechanism consistently outperformed baseline approaches in terms of solution quality. Additionally, tests on two credit classification datasets confirmed the method's effectiveness in improving search effectiveness beyond symbolic regression.
Implications
The findings suggest that incorporating feature importance into evolutionary algorithms can significantly enhance the performance of symbolic regression and classification tasks. This approach may lead to more robust models in various applications, particularly in domains where feature construction is critical, such as finance and data science.
HiRoute: Hierarchical Routed Prompt Tuning for Safety Alignment of Large Language Models
NLP
Large Language Models
Efficient ML
- HiRoute separates category-agnostic safety control from category-specific response guidance.
- The framework uses a hierarchical router to dynamically compose fine-grained prompts based on input risk.
- HiRoute maintains high safety rates while preserving the helpfulness of responses.
- The approach reduces over-refusal of benign inputs compared to traditional methods.
Read more
HiRoute: Hierarchical Routed Prompt Tuning for Safety Alignment of Large Language Models
Summary
The paper introduces HiRoute, a novel framework for safety alignment in large language models (LLMs) that addresses vulnerabilities to harmful requests and jailbreak attacks. Traditional prompt tuning methods often rely on static prompts that fail to adapt to the diverse nature of harmful inputs, leading to generic refusals that can degrade model performance. HiRoute overcomes these limitations by employing a hierarchical routing mechanism that separates category-agnostic safety control from category-specific response guidance. The framework first trains a lightweight hierarchical router to detect harmful intents and predict multi-label risk scores based on representations from a frozen LLM. Following this, preference optimization is used to learn a shared coarse-grained prompt alongside fine-grained prompt experts. During inference, benign inputs bypass the safety mechanism, while risky inputs utilize a combination of the shared prompt and router-weighted fine-grained prompts. The experimental results demonstrate that HiRoute achieves high safety rates across multiple benchmarks while maintaining helpfulness in responses and reducing over-refusal of benign inputs.
Methodology
HiRoute employs a two-stage training process. Initially, a hierarchical router is trained on representations from a frozen LLM to identify harmful intents and predict risk scores. Subsequently, both the backbone model and the router are frozen, and only the prompts are optimized using preference optimization. During inference, the framework dynamically constructs prompt combinations based on the routing results, allowing benign inputs to bypass safety prompting.
Results
HiRoute demonstrated high safety and helpfulness rates across various safety benchmarks, effectively balancing safety alignment with the need for constructive responses. The results indicate a significant reduction in over-refusal of benign inputs while maintaining competitive performance on general-purpose tasks.
Implications
The HiRoute framework has potential applications in enhancing the safety and reliability of LLMs in real-world scenarios, particularly in sensitive domains where harmful requests may arise. Its adaptive approach to prompt tuning could lead to more robust models capable of providing safe and helpful responses.
Forecast Collapse in Time-Series Foundation Models
Time Series
Optimization
Theory
- Forecast collapse is characterized by low forecast amplitude and poor cross-sectional structure in equity return predictions.
- The phenomenon is linked to target predictability, with trading volume forecasts performing better than equity returns.
- Two mechanisms are identified: low predictability constrains forecast amplitude, and per-series objectives overlook cross-series relationships.
- CalibRank, a new objective function, effectively balances calibration and ranking, improving forecast performance.
Read more
Forecast Collapse in Time-Series Foundation Models
Summary
This paper investigates a phenomenon termed 'forecast collapse' observed in time-series foundation models (TSFMs) when predicting hourly returns for 1,000 US equities. The authors find that predictions become nearly flat and exhibit poor stock ranking, which is measured by cross-sectional correlation. Interestingly, this collapse is not present when forecasting trading volume under the same conditions, suggesting a fundamental issue related to target predictability. The authors identify two primary mechanisms contributing to forecast collapse: low predictability limits the amplitude of calibrated point forecasts, while per-series objectives fail to capture cross-series structures. They introduce a new objective function, CalibRank, which balances calibration and ranking, significantly improving cross-sectional correlation while maintaining forecast amplitude. The findings highlight a critical blind spot in conventional time-series evaluation, emphasizing the need for metrics that consider cross-series structures essential for downstream decision-making.
Methodology
The authors conducted experiments using Finance1K, a dataset of hourly observations for US equities, to compare the predictability of returns and trading volumes. They analyzed the performance of twelve deep-learning forecasting models and derived theoretical relations to explain forecast collapse. The new objective, CalibRank, was tested across various models to assess its effectiveness in balancing calibration and ranking.
Results
The introduction of CalibRank nearly tripled the cross-sectional correlation while keeping the amplitude of forecasts close to the target. The results demonstrated that conventional per-series metrics could obscure failures in capturing necessary cross-series structures, which are critical for effective financial forecasting.
Implications
The findings suggest that financial forecasting models should be evaluated not only on per-series accuracy but also on their ability to maintain cross-series relationships. This could lead to improved decision-making in financial markets and a reevaluation of existing forecasting benchmarks.
When Local Variance Optimality Is Not Enough: RoPE-Aligned Q/K Rotations for Dynamic 4-Bit Quantisation
NLP
Large Language Models
Optimization
- Introduces a local benchmark for RoPE-commuting orthogonal transformations, focusing on independent pairwise rotations.
- Demonstrates that the head-shared pairwise configuration yields higher perplexity than full-head mixing in dynamic quantization settings.
- Establishes that the optimization of a surrogate does not guarantee improved quantization accuracy due to misalignment with quantizer statistics.
- Highlights the relationship between mixing support size and quantization error, suggesting broader mixing leads to better performance.
Read more
When Local Variance Optimality Is Not Enough: RoPE-Aligned Q/K Rotations for Dynamic 4-Bit Quantisation
Summary
This paper investigates the effectiveness of rotation-based post-training quantization methods in the context of transformer models, specifically focusing on RoPE (Rotary Position Embedding) aligned Q/K rotations. The authors challenge the conventional approach of applying orthogonal transforms across entire attention heads, proposing instead to respect the two-dimensional frequency pairs defined by RoPE. They derive the optimal rotation angle that minimizes channel variance under a pooled-covariance surrogate and verify its implementation. However, their experimental results reveal that this head-shared pairwise configuration does not enhance accuracy compared to full-head mixing, leading to increased perplexity across various checkpoints. The study highlights the misalignment between the surrogate's optimization and the quantizer's scale-setting statistic, suggesting that optimality in structured surrogates does not necessarily translate to reduced quantization error. The findings emphasize the importance of mixing support and the need for careful consideration of the quantization process in low-bit post-training quantization strategies.
Methodology
The authors derive the optimal rotation angle for minimizing channel variance using a pooled-covariance, position-averaged surrogate. They conduct experiments on various checkpoints, comparing the performance of head-shared pairwise configurations against full-head mixing in a dynamic quantization setting (W4A4KV4). The analysis includes evaluating perplexity and quantization error across different contexts and configurations.
Results
The experiments show that replacing the full-head Hadamard mixing with the head-shared pairwise configuration consistently increases perplexity across all evaluated contexts. Although the pairwise rotation can be optimized to attain a local minimum, it does not outperform the full-head mixing approach. Additionally, estimating the shared angle from K alone improves performance but still falls short of the full-head configuration.
Implications
The findings suggest that while structured surrogates can provide theoretical benefits, practical implementations must consider the alignment of optimization objectives with quantization processes. This has implications for the design of quantization strategies in transformer models, particularly in low-bit settings, where careful consideration of mixing support and quantization statistics is crucial.
Catching the Imposter: Self-Supervised Learning of Physical Coherence with Cross-Entity Feature Permutations
Time Series
- Introduces the 'imposter' pretext task for SSL that leverages physical coherence among features.
- Demonstrates the effectiveness of the imposter task on various downstream tasks in environmental modeling.
- Finds that the best pretext task depends on the specific downstream task family.
- Shows that combining the imposter task with existing SSL objectives enhances performance.
Read more
Catching the Imposter: Self-Supervised Learning of Physical Coherence with Cross-Entity Feature Permutations
Summary
This paper addresses the challenge of self-supervised learning (SSL) in scientific domains where features of entities are governed by physical laws. The authors introduce a novel pretext task called 'imposter,' which involves replacing subsets of an entity's features with those from another entity to train an encoder to identify these swapped features. This task encourages the model to learn cross-feature physical dependencies, as the swapped features are plausible in isolation but disrupt the physical coherence of the original entity. The authors evaluate their approach using global ERA5-Land reanalysis data across 21 environmental variables and assess the learned representations on seven downstream tasks, including climate classification and carbon flux estimation. They find that the effectiveness of the pretext task varies depending on the downstream task family and that the imposter task provides complementary information when combined with existing SSL objectives. This work highlights the importance of physical coherence as a source of self-supervision in scientific foundation models.
Methodology
The authors propose a discriminative pretext task where subsets of an entity's features are replaced with features from another entity. The model is trained to classify which features have been swapped, thereby learning the dependencies between features that maintain physical coherence. They also introduce a variant called 'multi-imposter' where each swapped feature comes from different donors, enhancing the challenge for the model.
Results
The study presents a systematic comparison of various self-supervised objectives for land-surface modeling, revealing that the imposter task is particularly effective when combined with other SSL methods. The results indicate that the choice of pretext task should align with the nature of the downstream tasks, suggesting that physical coherence is a significant factor in improving model performance.
Implications
This research suggests that incorporating physical coherence into SSL frameworks can lead to better representation learning in scientific domains, particularly in environmental sciences. It opens avenues for more effective modeling of complex systems where features are interdependent, potentially improving predictions in climate monitoring and resource management.
When Denoising Hurts: Rethinking the Terminal Step of Diffusion Time Series Forecasters -- Extended Version
Time Series
Generative Models
Optimization
- Identifies two distinct regimes in the reverse diffusion process: high-to-intermediate noise and low noise.
- Excessive low-noise refinement can introduce statistical drift and degrade forecast quality.
- Proposes an early-stopping mechanism that reduces sampling costs by 20-50% without requiring ground truth.
- Introduces a Bernoulli timestep-sampling strategy to emphasize high-noise predictions.
Read more
When Denoising Hurts: Rethinking the Terminal Step of Diffusion Time Series Forecasters -- Extended Version
Summary
This paper investigates the behavior of diffusion models in time series forecasting, challenging the assumption that iterative denoising always improves forecast quality. The authors identify two distinct regimes during the reverse diffusion process: a high-to-intermediate noise regime where the model recovers relevant temporal structures, and a low-noise regime where excessive denoising can lead to statistical drift and degrade forecast accuracy. To address these issues, the authors propose a label-free early-stopping mechanism that determines the optimal termination point of the reverse diffusion process, significantly reducing inference time and improving predictive accuracy. Additionally, they introduce a Bernoulli timestep-sampling strategy that focuses training on the high-noise region while maintaining coverage of the entire diffusion process. Their extensive experiments across eight real-world datasets demonstrate that their method outperforms existing approaches, achieving state-of-the-art probabilistic forecasting performance.
Methodology
The authors conducted a detailed analysis of the reverse diffusion process in time series forecasting, identifying the impact of noise levels on forecast quality. They developed an early-stopping mechanism based on the evolution of generated sample statistics and a Bernoulli timestep-sampling strategy to enhance training focus on high-noise regions. Experiments were performed on eight real-world datasets to validate their approach.
Results
The proposed methods demonstrated superior performance compared to existing diffusion-based forecasting approaches, achieving better accuracy while significantly reducing the computational cost associated with the sampling process. The early-stopping mechanism and Bernoulli sampling strategy were shown to effectively mitigate the issues of statistical drift and improve overall forecasting quality.
Implications
The findings suggest that careful consideration of the denoising process in diffusion models can lead to more efficient and accurate time series forecasting. The proposed strategies could be applied in various domains where time series data is prevalent, such as finance, weather forecasting, and supply chain management.
Resource-Adaptive Primal-Dual Learning for One-Warehouse Multi-Store Systems with Censored Demand
Optimization
Theory
- Introduction of Resource-Adaptive Primal-Dual Learning (RAPDL) for OWMS systems.
- RAPDL tracks the Primal-Dual re-solving path, adapting to changes in resource availability.
- Achieves logarithmic expected regret, outperforming existing square-root-order guarantees.
- Numerical experiments show RAPDL consistently lowers mean total costs compared to state-of-the-art methods.
Read more
Resource-Adaptive Primal-Dual Learning for One-Warehouse Multi-Store Systems with Censored Demand
Summary
This paper introduces Resource-Adaptive Primal-Dual Learning (RAPDL), a novel framework designed for one-warehouse multi-store (OWMS) systems facing censored demand. Traditional OWMS learning policies rely on fixed targets based on initial resource rates, which fail to adapt as sales change the available resources. RAPDL addresses this limitation by dynamically tracking the Primal-Dual re-solving path, allowing for real-time adjustments in store allocations and dual variables based on the evolving resource state. The framework utilizes remaining inventory and censored sales data to update allocation targets and resource states, leading to a logarithmic expected regret guarantee. This is a significant improvement over existing methods that only achieve square-root-order guarantees. The paper also provides a detailed analysis of the regret components and conducts numerical experiments demonstrating RAPDL's superior performance compared to established benchmarks across various inventory scenarios.
Methodology
The methodology involves developing a resource-adaptive framework that replaces fixed-target learning with online tracking of the Primal-Dual path. This is achieved by using remaining inventory and censored sales to inform updates to store allocations and dual variables, allowing for dynamic adjustments based on the evolving state of resources.
Results
The results indicate that RAPDL significantly reduces mean total costs in various inventory regimes when compared to the Double Binary Search policy, achieving lower costs in all 24 paired comparisons during numerical evaluations. The framework guarantees logarithmic regret, enhancing the performance of OWMS learning policies.
Implications
The implications of this research extend beyond OWMS systems, suggesting that the RAPDL framework could inform other online learning and control problems that involve depleting shared resources. This could enhance decision-making in various fields such as inventory management, logistics, and resource allocation.
The Objective Is the Bottleneck: Latent World Models Encode What Their Planners Cannot Use
Reinforcement Learning
Robotics
Optimization
- The predictor is not the bottleneck; it remains informative at long horizons.
- The planning objective can saturate and invert, affecting planning success.
- Long-horizon success is inversely related to one-step prediction accuracy.
- Reachability is a more effective objective than proximity for planning.
Read more
The Objective Is the Bottleneck: Latent World Models Encode What Their Planners Cannot Use
Summary
This paper investigates the limitations of latent world models in planning tasks, specifically focusing on the LeWorldModel in the TwoRoom environment. The author argues that the bottleneck in planning is not the predictor's capacity or horizon, but rather the planner's objective. Through experiments, it is shown that the predictor remains informative even at long horizons, while the planner is constrained to a shorter horizon. The study reveals that the planning objective can saturate and even invert, leading to counterintuitive behaviors where moving away from the goal can reduce the planner's cost. The paper also discusses the dissociation between prediction accuracy and long-horizon planning success, highlighting that a learned cost must be trained on the distribution relevant to the planner's scoring. The findings suggest that reachability, rather than proximity, should be the focus of planning objectives. The author provides a method to improve planning performance without retraining the model, demonstrating that the learned cost can outperform linear costs under certain conditions.
Methodology
The study employs a reproduction of the LeWorldModel on the TwoRoom environment, utilizing an 18.03M-parameter joint-embedding predictive architecture. The planning is conducted using cross-entropy-method search over learned dynamics, with specific configurations for candidate sequences and refinement iterations. The experiments are run on a standard CPU without any retraining of the model.
Results
The experiments reveal that the predictor's performance does not degrade significantly over long horizons, while the planner's effectiveness is limited to a shorter horizon. The objective function used in planning can lead to unexpected outcomes, such as lower costs when moving away from the goal. A simple change in the objective resulted in a significant increase in planning success rates, demonstrating that the learned cost can outperform traditional metrics under certain conditions.
Implications
The findings suggest that improving planning in latent world models may require a reevaluation of the objectives used in training. This could lead to more effective planning strategies in various applications, particularly in reinforcement learning and robotics, where long-term planning is crucial. The insights on reachability versus proximity could influence the design of future models and objectives.
Detecting Contaminated Code-Generation Prompt Batches via Influence Functions
Large Language Models
NLP
Generative Models
- CodeSIFT is a threat-model-agnostic method for detecting malicious code-generation prompts using influence functions.
- The method does not require prior knowledge of specific vulnerabilities, making it adaptable to novel attacks.
- Two benchmark datasets were created to evaluate the effectiveness of CodeSIFT against static analysis baselines.
- CodeSIFT achieved AUROC scores up to 0.98, outperforming traditional static analysis methods.
Read more
Detecting Contaminated Code-Generation Prompt Batches via Influence Functions
Summary
This paper addresses the vulnerabilities of large language models (LLMs) in code generation, particularly their susceptibility to prompts that can lead to insecure code implementations. Current defenses often rely on predefined threat models or known vulnerability patterns, which limits their effectiveness against novel attacks. The authors propose a new method called CodeSIFT, which utilizes influence functions to detect batches of prompts that may induce anomalous behavior in the model. Instead of identifying specific vulnerabilities, CodeSIFT assesses the influence of generated code on the model's parameters and employs a statistical test to determine if a set of prompts deviates from a benign reference distribution. The authors introduce two benchmark datasets, AuthSec and InfraCloud, to evaluate their approach against three static analysis baselines. The results demonstrate that CodeSIFT achieves an area under the receiver operating characteristic (AUROC) score of up to 0.98, significantly outperforming existing methods while maintaining low false positive rates. This suggests that influence-function-based detection could be a viable strategy for identifying malicious code-generation prompts without prior knowledge of specific attack classes.
Methodology
The authors employed influence functions to quantify the effect of individual prompts on the model's parameters. By measuring how the model's behavior deviates from a benign reference distribution, they developed a statistical test to identify contaminated prompt batches. This approach aggregates the influence signals across multiple prompts to assess the overall integrity of the batch.
Results
CodeSIFT demonstrated high effectiveness in detecting malicious prompts, achieving AUROC scores as high as 0.98 across three open-weight code LLMs with parameter sizes ranging from 3B to 7B. The method maintained well-calibrated false positive rates and significantly outperformed static analysis baselines such as Bandit and Semgrep.
Implications
The findings suggest that influence-function-based detection methods could enhance the security of code generation systems by identifying potentially harmful prompts without relying on predefined vulnerability patterns. This could lead to more robust coding assistants and safer software development practices.
Boosting Data Augmentation with Stochastic Weight Averaging
Computer Vision
Graph Learning
Theory
- Combines data augmentation with Stochastic Weight Averaging (SWA) to enhance model performance.
- Introduces a theoretical framework using OrnsteinβUhlenbeck processes to analyze the effects of SWA.
- Demonstrates that the combination of SWA and data augmentation leads to significant equivariance improvements.
- Validates theoretical findings through extensive numerical experiments on diverse datasets and model architectures.
Read more
Boosting Data Augmentation with Stochastic Weight Averaging
Summary
This paper investigates the integration of data augmentation with Stochastic Weight Averaging (SWA) to enhance the equivariance properties of neural networks. The authors highlight that while data augmentation is a flexible method for incorporating symmetries into deep learning models, traditional ensemble methods require extensive computational resources due to repeated training. They propose using SWA, which averages weights from a single training trajectory, as a more efficient alternative. By modeling the stochastic training trajectory with an OrnsteinβUhlenbeck process, the authors derive theoretical insights into the performance boost provided by SWA when combined with augmented data. They demonstrate that this combination yields an equivariance boost that surpasses the performance improvements expected from SWA alone. The paper includes extensive numerical experiments across various models in computer vision and graph classification, confirming the theoretical predictions and showcasing the benefits of this approach in practical scenarios.
Methodology
The authors analyze SWA by approximating the stochastic training trajectory with an OrnsteinβUhlenbeck process. They quantify performance and equivariance boosts using loss ratios and derive theoretical bounds in the infinite-width limit using neural tangent kernels (NTKs). Numerical experiments validate these theoretical results across various models and tasks.
Results
The study finds that SWA combined with data augmentation leads to a non-trivial improvement in equivariance, supported by theoretical bounds and extensive numerical validation. The results show consistent performance gains across multiple datasets, including computer vision and graph classification tasks.
Implications
This work suggests that integrating SWA with data augmentation can significantly enhance the performance of neural networks, particularly in tasks where symmetries play a crucial role. It opens avenues for more efficient training methods that leverage the benefits of both techniques without the computational burden of traditional ensemble methods.
Designing Compact Neural Architectures via Neuron Gating and Mixed Activation
Optimization
Efficient ML
Theory
- Introduces a scalable method for Neural Architecture Search (NAS) using continuous relaxations.
- Develops three bilevel optimization formulations applicable to various neural network types.
- Demonstrates that the proposed methods outperform existing NAS techniques like DARTS.
- Achieves high accuracy with significantly fewer parameters in the identified architectures.
Read more
Designing Compact Neural Architectures via Neuron Gating and Mixed Activation
Summary
This paper addresses the challenges of Neural Architecture Search (NAS), which is often computationally expensive due to the discrete nature of architectural decisions and the high cost of training candidate architectures. The authors propose a novel approach that formulates NAS as a bilevel optimization problem, allowing for continuous relaxations of neuron and activation-level decisions. This results in differentiable optimization methods that can be applied across various neural network architectures, including MLPs, CNNs, RNNs, and Transformers. The proposed methodsβNAS based on Neuron Gating (NAS-NG), NAS based on Mixed Activation (NAS-MA), and NAS based on Neuron Gating and Mixed Activation (NAS-NGMA)βare evaluated on the MNIST and CIFAR-10 datasets. The results demonstrate that these methods consistently identify compact architectures with competitive or improved predictive performance compared to existing methods like Differentiable Architecture Search (DARTS). Notably, NAS-NGMA achieves 98.68% accuracy on MNIST with 7.69M parameters, while NAS-NG reaches 99.63% accuracy with only 0.26M parameters. The findings suggest that relaxed bilevel optimization can serve as an effective alternative to discrete search methods in NAS, enabling the discovery of high-performing architectures with fewer parameters.
Methodology
The authors propose three methods for NAS that utilize continuous relaxations of neuron and activation decisions, allowing for differentiable optimization. These methods are formulated as bilevel optimization problems and solved using hypergradient-based techniques, enabling efficient exploration of the architecture space.
Results
The proposed methods were tested on MLP and CNN architectures using the MNIST and CIFAR-10 datasets. NAS-NGMA achieved 98.68% test accuracy with 7.69M parameters on MNIST, while NAS-NG reached 99.63% accuracy with only 0.26M parameters. On CIFAR-10, the proposed methods consistently outperformed DARTS, demonstrating their effectiveness in identifying compact architectures with strong predictive performance.
Implications
The findings suggest that the proposed bilevel optimization framework can significantly reduce the computational burden associated with NAS, making it a practical approach for discovering efficient neural architectures. This could lead to advancements in various applications of deep learning, particularly in resource-constrained environments.
When Does More Correct Data Hurt? Insertion-Stability and the Limits of Dimension-Based Theory
Theory
- Insertion-stability is a key property that determines whether adding correct data will hurt or help a learner's performance.
- The paper establishes that no learner can achieve better than Ξ(d/n) log(en/d) expected error for classes with VC dimension d β₯ 2.
- Closure is shown to be insertion-stable, allowing intersection-closed classes to maintain their clean error rates despite data augmentation.
- The relationship between the learner and the class is critical; a mismatch can lead to increased error rates even with correct data.
Read more
When Does More Correct Data Hurt? Insertion-Stability and the Limits of Dimension-Based Theory
Summary
This paper investigates the counterintuitive phenomenon where adding correctly labeled data to a training set can actually increase the expected error of a learning algorithm. The author builds on the work of Larsen, Pabbaraju, and Shetty, who modeled the problem using a monotone adversary that can append additional examples to a training set, provided they are correctly labeled. The study reveals that for classes with VC dimension d β₯ 2, no learner can guarantee an expected error better than a logarithmic factor above the clean PAC rate. The paper introduces the concept of 'insertion-stability,' where a learner's error region shrinks with the addition of correct examples, ensuring that the risk after insertions does not exceed the risk on the clean part alone. The author proves that certain classes retain their clean error rates under arbitrary insertions, while others do not, emphasizing that the interaction between the learner and the class is crucial. The findings challenge traditional dimension-based theories, suggesting that the geometry of the learner's error region is more significant than classical dimensions in determining the impact of data augmentation.
Methodology
The paper employs theoretical analysis to explore the concept of insertion-stability and its implications for learning algorithms. It builds on existing models of adversarial data insertion and derives new results regarding the interaction between learners and data classes. The author uses combinatorial techniques and reductions to prove the main results, including the insertion-stability lemma and the properties of intersection-closed classes.
Results
The main results include the establishment of the insertion-stability lemma, which shows that if a learner's error region shrinks with correct insertions, its adversarial risk is dominated by its clean risk. The paper also demonstrates that closure is insertion-stable, retaining clean rates under adaptive insertions. Additionally, it reveals that certain classes, despite having the same VC and Littlestone dimensions, can exhibit different rates of error under monotone insertions, highlighting the importance of the learner-class pairing.
Implications
The findings suggest that practitioners should be cautious when augmenting datasets, as adding correct data can sometimes lead to worse performance. The concept of insertion-stability could guide the design of more robust learning algorithms that can effectively handle data augmentation without incurring additional error. This research also calls for a reevaluation of dimension-based theories in machine learning, emphasizing the need for a deeper understanding of the learner's error geometry.
Neural Quadratic Forms: A Unified Minimal Model for Sudden Learning and Scaling Laws
Theory
Optimization
- Introduces Neural Quadratic Forms (NQF) as a minimal model for understanding sudden learning and scaling laws in neural networks.
- Demonstrates that neural architectures exhibit a symmetry that allows for a unified representation of different models through a structure matrix A(x).
- Establishes a connection between training dynamics and collective variables, leading to predictions about feature acquisition timing and power law behaviors.
- Validates the theoretical predictions through numerical experiments across diverse neural network architectures and training methodologies.
Read more
Neural Quadratic Forms: A Unified Minimal Model for Sudden Learning and Scaling Laws
Summary
This paper introduces a unified minimal model termed Neural Quadratic Forms (NQF) that explains the abrupt learning dynamics observed in neural networks trained via gradient descent, alongside the smooth power laws governing training losses across various architectures. The authors leverage the symmetry inherent in neural network architectures, which allows for the characterization of training dynamics through a universal quadratic form. By establishing that neural network layers can be represented as sums of interchangeable components, the authors derive a structure matrix A(x) that encapsulates architectural details while maintaining a common framework for different models, including perceptrons, attention layers, and convolutions. The training dynamics are shown to converge on an order parameter M = WWβΊ, leading to a Lotka-Volterra equation that describes the sequential activation of modes during training. This model successfully predicts the timing of feature acquisition and the emergence of power laws in training time, validated through numerical experiments across various architectures and training methods.
Methodology
The authors employ a theoretical framework based on symmetry principles to derive a universal quadratic form for neural network training dynamics. They analyze the implications of this form across various architectures, focusing on the role of the structure matrix A(x) and the order parameter M = WWβΊ. Numerical simulations are conducted to validate the theoretical predictions regarding learning dynamics and scaling laws.
Results
The study confirms that the derived NQF accurately captures the abrupt learning phases and smooth power law behaviors observed in neural network training. The model predicts that the timing of feature acquisition is influenced by the scale of initial weights, with smaller weights leading to more pronounced plateaus in training dynamics. The theoretical framework aligns with empirical observations across multiple architectures, demonstrating the universality of the model.
Implications
The findings suggest that understanding the collective dynamics of neural networks can inform the design of more efficient architectures and training strategies. The NQF model provides a theoretical basis for predicting learning behaviors, which could enhance the scalability and performance of neural networks in various applications.
Latent On-Policy Self-Distillation
Reinforcement Learning
Large Language Models
Robotics
- Introduces a learnable privileged context for on-policy self-distillation.
- Demonstrates superior performance over existing OPSD methods.
- Achieves high learning efficiency with reduced rollout budget.
- Employs a privileged-margin objective to enhance teacher-student dynamics.
Read more
Latent On-Policy Self-Distillation
Summary
This paper introduces Latent On-Policy Self-Distillation (LOPD), a novel framework aimed at enhancing the learning process of agents by allowing them to learn from their own experiences without relying on pre-defined privileged contexts. Traditional on-policy self-distillation methods depend on designer-specified artifacts for supervision, which limits their scalability and adaptability. LOPD addresses this by making the privileged context learnable and end-to-end, allowing the self-teacher to derive relevant experiences from the agent's interaction history. The framework retrieves experiences and composes them into continuous latent tokens that condition the self-teacher, which then provides dense token-level supervision to the student. A privileged-margin objective is introduced to ensure that the teacher maintains an advantage over the student during training. Empirical results show that LOPD outperforms existing methods in both agentic tool use and code generation tasks while demonstrating high learning efficiency, requiring less than 30% of the rollout budget compared to other methods. The findings suggest that making the privileged context learnable is crucial for achieving significant performance gains, positioning LOPD as a scalable solution for self-directed agent evolution.
Methodology
LOPD utilizes a framework where the student agent generates trajectories from its task and interaction history. A self-teacher, conditioned on learnable latent tokens derived from relevant experiences, provides dense supervision. The training process employs reverse-KL distillation, and a privileged-margin constraint ensures the teacher maintains a verifiable advantage over the student.
Results
LOPD outperformed traditional methods such as OPSD, RLVR, and Skill-SD in various tasks, including agentic tool use and code generation. It demonstrated high learning efficiency, achieving results with less than 30% of the rollout budget compared to competitors. Ablation studies confirmed the necessity of learnable privileged context for performance improvements.
Implications
The findings suggest that LOPD can lead to more scalable and adaptable AI systems capable of self-improvement through experience. This approach could be applied in various domains requiring continual learning and adaptation, such as robotics, autonomous systems, and interactive AI applications.
Structure-preserving uncertainty quantification for GENERIC dynamics
Theory
- Introduction of S-PENNs for uncertainty quantification in hard-constrained models.
- Preservation of thermodynamic consistency in stochastic realizations.
- Significant reduction in computational costs compared to deep ensembles.
- Validation through multiple numerical examples demonstrating effectiveness.
Read more
Structure-preserving uncertainty quantification for GENERIC dynamics
Summary
This paper introduces Structure-Preserving Epistemic Neural Networks (S-PENNs), a novel framework for uncertainty quantification (UQ) in scientific machine learning models that adhere to hard architectural constraints. The authors focus on the GENERIC (General Equation for Non-Equilibrium Reversible-Irreversible Coupling) dynamics, which is essential for modeling systems governed by both reversible and irreversible processes. S-PENNs maintain the structural integrity of a pretrained model by integrating lightweight epinets into its constrained components, ensuring that all sampled realizations are physically admissible. The framework is combined with split conformal prediction to generate prediction intervals with finite-sample coverage guarantees. The authors validate S-PENNs through three numerical examples: a harmonic oscillator coupled to a heat bath, an idealized chemical motor, and a one-dimensional viscoplastic model. The results demonstrate that S-PENNs yield thermodynamically consistent stochastic realizations and well-calibrated prediction intervals while significantly reducing computational costs compared to traditional deep ensembles. Although the study is centered on GENERIC dynamics, the proposed framework has broader applicability in scientific machine learning models with hard or soft constraints.
Methodology
The authors propose S-PENNs, which integrate lightweight epinets into existing constrained models to ensure physical admissibility. They also employ split conformal prediction for post-hoc calibration of prediction intervals. The methodology is validated through numerical simulations of various dynamical systems governed by ordinary and partial differential equations.
Results
S-PENNs produced thermodynamically consistent stochastic realizations and well-calibrated prediction intervals across three numerical examples, achieving a computational efficiency improvement of 1 to 3 orders of magnitude compared to deep ensembles.
Implications
The proposed framework can enhance the reliability of predictions in scientific machine learning applications, particularly in fields requiring adherence to physical laws. It opens avenues for more efficient modeling of complex systems in computational mechanics and beyond.
L-FNO: Lorentzian Fourier Neural Operator for Stochastic Event Dynamics
Time Series
- L-FNO incorporates a Lorentzian spectral kernel to model self-exciting dynamics in stochastic events.
- The model is trained using a Poisson negative log-likelihood objective, improving calibration and sensitivity to rare events.
- L-FNO outperforms traditional regression-based models in event likelihood and rare-event detection across multiple benchmarks.
- The learned parameters provide interpretable insights into the predictability structure of stochastic event systems.
Read more
L-FNO: Lorentzian Fourier Neural Operator for Stochastic Event Dynamics
Summary
The paper introduces the Lorentzian Fourier Neural Operator (L-FNO), a novel stochastic neural operator designed to model stochastic event dynamics, particularly in scenarios characterized by rare and bursty events. Traditional neural operators, which typically function as regression models, struggle with sparse event data due to their inability to capture the conditional intensity of events. L-FNO addresses this gap by integrating a Lorentzian spectral kernel into the Fourier Neural Operator framework, allowing for the modeling of history-dependent excitation akin to Hawkes processes. The authors evaluate L-FNO on eight synthetic benchmarks and three real-world datasets, including applications in disease outbreak prediction and semiconductor defect detection. The results demonstrate that L-FNO significantly enhances event likelihood, calibration diagnostics, and rare-event detection compared to existing regression- and likelihood-based neural operator models. The findings suggest that structured spectral memory and likelihood-based learning provide effective inductive biases for modeling stochastic event dynamics, moving beyond mere prediction to quantifying the temporal structure of events.
Methodology
The L-FNO extends the Fourier Neural Operator architecture by integrating a Lorentzian spectral kernel, which corresponds to an exponential excitation function. This allows the model to learn an event intensity function that captures both covariate-driven effects and history-dependent dynamics. The model is trained using a Poisson negative log-likelihood loss to address the limitations of traditional regression approaches in sparse data scenarios.
Results
L-FNO was evaluated on eight synthetic point-process benchmarks and three real-world datasets, showing improvements in event likelihood, calibration diagnostics, and rare-event detection compared to regression- and likelihood-based neural operator baselines. The model effectively captures the temporal structure of stochastic events, demonstrating its capability to learn from sparse data.
Implications
The findings suggest that L-FNO can be applied in various fields where understanding and predicting rare events are crucial, such as epidemiology, finance, and manufacturing. The model's ability to quantify the temporal structure of events can enhance decision-making processes in operational systems facing uncertainty.
Quantum Multi-Armed Bandits and Linear Bandits: Lower Bounds and Algorithms
Theory
- Establishes minimax lower bounds for quantum multi-armed and linear bandits.
- Proves that logarithmic regret dependence on T is unavoidable for QMAB.
- Introduces a design-based elimination algorithm that reduces regret dependence on dimension.
- Implements a low-bias, low-variance quantum mean estimator to optimize performance.
Read more
Quantum Multi-Armed Bandits and Linear Bandits: Lower Bounds and Algorithms
Summary
This paper investigates quantum multi-armed bandits (QMAB) and quantum linear bandits (QLB) within the framework established by Wan et al. (2023). The authors address two critical open questions regarding the regret bounds of these quantum bandit models. They establish the first minimax lower bounds for QMAB and finite-action QLB, proving that the regret scales as β¦(K log(T/K)) for QMAB and β¦(d log(T/d)) for finite-action QLB, thus confirming that logarithmic dependence on T is unavoidable. The authors employ a high-confidence single-arm quantum testing lower bound to derive these results, utilizing the polynomial method and a Remez-type inequality. Additionally, they propose a design-based elimination algorithm for finite-action QLB, which significantly improves the regret bound from O(dΒ² polylog T) to nearly linear in d, specifically O(d log(KT) Β· polylog(d, T)), when the action set size is polynomial in d. This algorithm combines a low-bias, low-variance quantum mean estimator with a G-optimal design, optimizing query allocation based on design weights. The findings not only provide theoretical insights into quantum bandit problems but also enhance algorithmic performance in practical applications.
Methodology
The authors utilize a combination of theoretical proofs and algorithm design. They prove lower bounds using a reduction from bandit regret to a point-versus-interval testing problem, employing polynomial methods and inequalities. For the algorithm, they integrate a low-bias quantum mean estimator with a G-optimal design to optimize query allocation and reduce regret.
Results
The paper presents two main results: a minimax lower bound of β¦(K log(T/K)) for QMAB and β¦(d log(T/d)) for finite-action QLB, confirming the intrinsic logarithmic dependence on T. Additionally, the proposed LV-G-Elim algorithm achieves expected regret of O(d log(KT) Β· polylog(d, T)), significantly improving upon previous bounds.
Implications
The findings have significant implications for the development of quantum algorithms in bandit settings, enhancing both theoretical understanding and practical performance. The results may influence future research in quantum computing applications, particularly in areas requiring efficient decision-making under uncertainty.
Approximate Muon with low-rank adapters
Optimization
Efficient ML
- sMuon provides a mathematically efficient method for training LoRA adapters with Muon, requiring less optimizer state than AdamW.
- The proposed method is implemented using only matrix multiplications, making it faster and more memory-efficient than existing approaches.
- sMuon shows strong performance across multiple fine-tuning benchmarks, particularly with Muon-pretrained models.
- The method achieves competitive results in ReLoRA pretraining, demonstrating advantages over AdamW-based methods.
Read more
Approximate Muon with low-rank adapters
Summary
This paper addresses the challenges of using the Muon optimizer for Parameter-Efficient Fine-Tuning (PEFT) of neural networks, particularly in conjunction with the popular LoRA method. The authors propose a new approach, sMuon, which approximates the Muon objective in a low-rank setting through linearization and least-squares techniques. This method allows for efficient implementation using only matrix multiplication operations, avoiding the complexity of traditional linear algebra decompositions. The authors demonstrate that sMuon achieves competitive performance across various benchmarks, particularly excelling in scenarios involving Muon-pretrained models. The findings suggest that using Muon for low-rank fine-tuning can yield moderate performance improvements, thereby enhancing its applicability in the field of neural network optimization.
Methodology
The authors linearize the low-rank Muon objective to derive an optimal descent step within the parameter space. They introduce a matmul-only algorithm, sMuon, which efficiently computes the necessary updates without resorting to expensive matrix decompositions. The method is validated through a series of experiments comparing its performance against other low-rank Muon optimizers.
Results
sMuon outperforms existing methods in several benchmarks, achieving the highest accuracy in six out of eleven tasks when fine-tuning a Muon-pretrained model. In the ReLoRA pretraining context, sMuon achieves the lowest final validation loss while being approximately 30% faster than Riemannion.
Implications
The findings suggest that sMuon can enhance the efficiency and effectiveness of fine-tuning large neural networks, particularly in scenarios where computational resources are limited. This could lead to broader adoption of Muon in practical applications of machine learning.
Learning Discrete Decisions for MIPs with Constraint-Aware Diffusion
Optimization
Generative Models
Graph Learning
- Introduction of Constrained Graph Diffusion (CGD) for MIPs, which enforces feasibility during the decision-making process.
- Decomposition of mixed-integer nonlinear programming (MINLP) problems into discrete and continuous components to improve optimization efficiency.
- Demonstrated effectiveness of CGD on diverse applications, achieving substantial improvements in solution quality and feasibility.
- Achieved significant computational speedups compared to traditional numerical solvers.
Read more
Learning Discrete Decisions for MIPs with Constraint-Aware Diffusion
Summary
This paper presents a novel learning-based approach to tackle mixed-integer optimization problems (MIPs) using a method called Constrained Graph Diffusion (CGD). MIPs are challenging due to the need to simultaneously determine discrete and continuous variables while adhering to complex constraints. The CGD framework employs a graph-based generative diffusion model that learns the discrete decisions of MIPs and integrates a training-free feasibility projection operator into the reverse diffusion process. This integration ensures that intermediate samples remain within the feasible set during generation. Once the discrete decisions are established, the problem simplifies to a continuous optimization task that can be efficiently solved using existing numerical methods. The authors demonstrate the effectiveness of CGD on two distinct applications: optimal transmission switching for AC optimal power flow and mixed-integer portfolio optimization. The results show that CGD significantly enhances feasibility and solution quality compared to existing learning-based methods and achieves speedups of up to 425 times over state-of-the-art numerical solvers for mixed-integer nonlinear programs (MINLPs).
Methodology
The methodology involves a graph-based generative diffusion model that learns the distribution of discrete decisions while incorporating feasibility projections during the reverse diffusion process. This allows for the generation of feasible discrete decisions, which are then fixed to solve the remaining continuous optimization problem using standard numerical methods.
Results
The CGD framework was evaluated on optimal transmission switching for AC optimal power flow and mixed-integer portfolio optimization, showing improved feasibility and solution quality over existing learning-based approaches. The method also achieved speedups of up to 425 times compared to state-of-the-art numerical solvers for MINLPs.
Implications
The proposed CGD framework has the potential to enhance the efficiency and effectiveness of solving mixed-integer optimization problems across various domains, including power systems, finance, and logistics, where decision-making involves both discrete and continuous variables.
Connected Subspace Clustering: Hardness, a Scalable Heuristic, and an Application to Sea Level Geodesy
Optimization
Time Series
Theory
- Introduction of Connected Subspace Clustering, combining connectivity and subspace reconstruction.
- Proven NP-hardness of the problem, even for simple cases.
- Development of an efficient heuristic that ensures connected clusters.
- Strong performance in sea level geodesy applications, outperforming traditional clustering methods.
Read more
Connected Subspace Clustering: Hardness, a Scalable Heuristic, and an Application to Sea Level Geodesy
Summary
This paper introduces the Connected Subspace Clustering problem, which integrates the concepts of connectivity and subspace clustering to address the challenge of grouping high-dimensional data while maintaining physical coherence. The authors demonstrate that this problem is NP-hard to approximate, even in simplified cases, highlighting its computational complexity. To tackle this, they propose a Lloyd-style heuristic that alternates between fitting subspaces and merging components to ensure connectivity. The method is evaluated in the context of sea level geodesy, where it is crucial to identify contiguous regions of sea surface measurements for further analysis. Experimental results show that the proposed heuristic outperforms existing methods, achieving connected clusters with lower reconstruction costs and better alignment with climate indices. The approach is not limited to geodesy but can be applied to various spatially embedded multivariate time series, making it broadly relevant across multiple domains.
Methodology
The authors present a heuristic that alternates between subspace fitting and an iterative component-merging procedure to enforce connectivity. This method is designed to work with a given connectivity graph and aims to minimize the total squared distance of points to their best-fit affine subspaces.
Results
In extensive experiments involving 160 configurations of global sea level time series, the proposed method achieved connected regions in 73.75% of configurations, significantly reducing the number of disconnected fragments compared to unconstrained methods. The results also aligned well with established climate indices, indicating the effectiveness of the clustering approach.
Implications
The Connected Subspace Clustering approach has significant implications for analyzing spatially embedded multivariate time series data, particularly in fields such as climate science, remote sensing, and neuro-imaging. Its ability to maintain connectivity while clustering can enhance the interpretability and accuracy of analyses in these domains.
Defensive Boosting for Online Probabilistic Forecasting
Theory
Efficient ML
Optimization
- The Defensive Booster algorithm achieves both competitive Brier scores and low classification error under varying conditions.
- It operates efficiently with only one weak-class learner, unlike previous methods that require multiple learners.
- The algorithm provides local hard-core certificates, enhancing its adaptability to different time intervals.
- Experimental results show substantial improvements in predictive performance and runtime efficiency over existing methods.
Read more
Defensive Boosting for Online Probabilistic Forecasting
Summary
This paper addresses the challenge of online probabilistic forecasting of binary outcomes in the presence of an adaptive adversary. The authors propose a novel algorithm called the Defensive Booster, which combines the strengths of existing online boosting techniques to provide two key guarantees: competitive Brier scores with the best predictor from a weak hypothesis class and the ability to drive classification error to zero under a weak-learning condition. The Defensive Booster operates efficiently by utilizing a single weak-class learner, contrasting with prior methods that require maintaining large ensembles of weak learners. The paper also introduces a strongly adaptive variant of the algorithm that provides local hard-core certificates on every time interval, ensuring robust performance across varying conditions. Experimental results demonstrate that the Defensive Booster not only outperforms existing baselines in terms of predictive accuracy but also achieves significantly faster runtimes, making it a promising approach for real-time forecasting tasks.
Methodology
The Defensive Booster is developed as a black-box reduction from an online learning algorithm for a weak hypothesis class. It operationalizes the dual view of boosting, utilizing mistake weights to certify the failure of the weak-learning condition when errors persist. The algorithm is designed to be efficient, requiring only one weak learner while maintaining strong performance guarantees.
Results
The Defensive Booster demonstrates competitive Brier scores with the best predictor in the span of weak hypotheses and achieves low classification error when the smooth weak-learning condition is satisfied. The algorithm's performance is validated through experiments on synthetic and real data streams, showing improvements over prior baselines in both accuracy and runtime.
Implications
The findings suggest that the Defensive Booster can be effectively applied in scenarios requiring real-time probabilistic forecasting, such as financial predictions, risk assessment, and adaptive decision-making systems. Its efficiency and strong performance guarantees make it a valuable tool in online learning contexts.
Generating Benchmark Health Data Using a Tabular Diffusion Transformer
Generative Models
- Introduces a two-stage framework for generating synthetic tabular data from multiple heterogeneous tables.
- Transforms raw tables into standardized statistical tables to capture essential statistical properties.
- Utilizes a diffusion transformer model to learn structural patterns and generate synthetic data.
- Demonstrates high fidelity in statistical representations and effective diversity in generated datasets.
Read more
Generating Benchmark Health Data Using a Tabular Diffusion Transformer
Summary
This paper addresses the limitations of existing synthetic tabular data generation (STDG) methods, which typically focus on single-input-table scenarios and struggle with multiple heterogeneous tables. The authors propose a two-stage framework for Cross-Tabular Data Generation (CTDG). In the first stage, heterogeneous raw tables are transformed into standardized statistical tables, capturing marginal distributions and pairwise correlations. The second stage involves training a diffusion transformer model to learn structural patterns from these homogeneous statistical tables, enabling the generation of synthetic statistical tables. The synthetic raw tables are reconstructed through multivariate Gaussian sampling and inverse probability integral transform. The proposed framework allows for the generation of an unlimited number of realistic synthetic heterogeneous tables, demonstrating high fidelity in learned statistical representations and a favorable fidelity-diversity trade-off in the generated data. This approach is particularly valuable in healthcare for creating benchmark datasets that reflect real-world data distributions and support privacy-preserving data sharing.
Methodology
The methodology involves a two-stage process: first, transforming heterogeneous raw tables into standardized statistical tables that capture marginal distributions and correlations; second, training a diffusion transformer model on these tables to generate synthetic statistical tables, which are then reconstructed into raw tables using multivariate Gaussian sampling and inverse probability integral transform.
Results
The experimental results indicate that the proposed CTDG framework achieves high fidelity in the learned statistical representations and maintains a favorable trade-off between fidelity and diversity in the synthetic data generated, validating the effectiveness of the approach.
Implications
The implications of this work include the potential for creating diverse and realistic benchmark datasets for machine learning research, particularly in healthcare, where synthetic patient records can facilitate data sharing, mitigate data scarcity, and enhance the robustness of clinical machine learning pipelines.
Stochastic Control Policies for Robust Molecular Transition Path Sampling
Reinforcement Learning
Optimization
Theory
- Introduction of two stochastic control policies (FS-TPS and LaS-TPS) for molecular transition path sampling.
- FS-TPS improves transition success and path quality by utilizing a state-dependent Gaussian distribution.
- LaS-TPS enhances robustness by sampling from a latent variable representation, capturing correlated atomic motions.
- Both methods show significant improvements over deterministic baselines across multiple biomolecular systems.
Read more
Stochastic Control Policies for Robust Molecular Transition Path Sampling
Summary
This paper addresses the challenge of efficiently generating rare molecular transition trajectories between metastable states, which is crucial for understanding biomolecular mechanisms. The authors propose two stochastic control policies, FS-TPS and LaS-TPS, to enhance the robustness and exploration capabilities of transition path sampling (TPS) methods. FS-TPS introduces a state-dependent Gaussian distribution for control policy outputs, while LaS-TPS employs a latent variable approach to capture correlated atomic motions. Extensive experiments on biomolecular systems of varying sizes (alanine dipeptide, chignolin, and BBL) demonstrate that these stochastic policies significantly improve transition success rates and path quality compared to deterministic baselines, while also reducing sensitivity to random initialization. The findings suggest that incorporating stochasticity into policy learning can lead to more reliable and effective sampling of molecular transitions.
Methodology
The authors recast rollout-based control as learning a path-space proposal distribution, introducing stochasticity into the control policies. FS-TPS directly parameterizes a state-dependent Gaussian distribution for control outputs, while LaS-TPS samples from a compact latent variable and decodes it into structured force variations. The methods are evaluated through multi-seed experiments on various biomolecular systems.
Results
FS-TPS achieved improvements of 39.5% in RMSD, 52.8% in transition success, and 22.2% in path diversity on alanine dipeptide compared to deterministic methods. LaS-TPS further reduced RMSD and transition-state energy by 41.9% and 25.4%, respectively, while improving transition success and mode coverage by 58.9% and 77.8%. Both methods maintained their advantages across larger systems like chignolin and BBL.
Implications
The findings suggest that incorporating stochastic control policies can enhance the efficiency and reliability of molecular transition path sampling, which has implications for studying complex biomolecular processes such as protein folding and conformational changes. This approach could lead to better understanding and prediction of biomolecular behavior in various scientific and industrial applications.
HI-MeshGraphNets: Efficient and Accurate Mesh-based Physics Learning with Hierarchical Multi-scale Graph Neural Networks
Graph Learning
Efficient ML
Theory
- HI-MGN enhances long-range communication through hierarchical message passing.
- Utilizes FPS-Voronoi coarsening to create coarse graphs from original mesh adjacency.
- Incorporates learned interpolation for reconstructing fine-scale features post-coarsening.
- Achieves improved accuracy while reducing memory and training costs.
Read more
HI-MeshGraphNets: Efficient and Accurate Mesh-based Physics Learning with Hierarchical Multi-scale Graph Neural Networks
Summary
The paper introduces Hierarchical Interpolating MeshGraphNets (HI-MGN), a novel approach to enhance the efficiency and accuracy of mesh-based physics learning using hierarchical multi-scale graph neural networks (GNNs). Traditional GNNs face challenges with long-range interactions in large, high-fidelity meshes due to flat message passing, which can lead to increased computational costs and memory usage. HI-MGN addresses this by implementing a hierarchical message passing mechanism that allows for improved long-range communication. The method involves coarsening the original mesh into coarse graphs using farthest-point sampling (FPS) and Voronoi partitioning, which preserves the mesh topology while enabling information to propagate over larger distances with fewer steps. After processing on the coarse graphs, a learned interpolation network reconstructs fine-scale features back to the original mesh resolution. The authors demonstrate the effectiveness of HI-MGN through numerical comparisons on structural and fluid problems, showing that it achieves higher accuracy and reduced training time and memory usage compared to existing GNN methods like MeshGraphNets and Bi-Stride Multi-Scale GNN (BSMS-GNN). This work highlights the potential of topology-aware hierarchical processing for scalable mesh-based physics surrogate modeling.
Methodology
The methodology involves a hierarchical multi-scale GNN architecture that coarsens the original mesh into coarse graphs using farthest-point sampling and Voronoi partitioning. This allows for efficient long-range communication during message passing, followed by a learned interpolation network that reconstructs the original mesh features.
Results
HI-MGN outperforms existing GNN methods in terms of accuracy while maintaining a comparable number of message passing blocks. It also significantly reduces training time and peak memory usage, demonstrating its efficiency in handling large unstructured meshes.
Implications
The findings suggest that HI-MGN can be effectively utilized for scalable mesh-based physics surrogate modeling, potentially benefiting applications in computational fluid dynamics, structural analysis, and other fields requiring high-fidelity simulations.
Scaling Automatic Research Agents via World Models
Reinforcement Learning
Large Language Models
Efficient ML
- Introduction of World Model RL (WMRL) to replace expensive environment execution in AutoResearch agents.
- Development of Online Debiasing and Inverse-Variance Denoising mechanisms to enhance the performance of the world model.
- Theoretical grounding of the framework, proving improved convergence guarantees.
- Empirical validation showing 3-4x acceleration in training and superior performance compared to larger models.
Read more
Scaling Automatic Research Agents via World Models
Summary
This paper addresses the challenge of scaling Automatic Research (AutoResearch) agents, which utilize large language models (LLMs) to conduct empirical research autonomously. The authors identify a critical bottleneck in the reinforcement learning (RL) training process, where the environment execution component of agent trajectories incurs significantly higher costs compared to the agent generation component. To mitigate this issue, they propose a novel approach called World Model RL (WMRL), which substitutes real environment execution with a world model that simulates execution outcomes. This allows for more efficient training by enabling the batching of trajectories, thus alleviating the execution cost bottleneck. The authors also introduce two correction mechanismsβOnline Debiasing and Inverse-Variance Denoisingβto address biases and noise in the world model's outputs. Theoretical proofs demonstrate that these mechanisms improve convergence guarantees. Empirical results show that WMRL accelerates training by 3-4 times across various tasks and scales, outperforming larger baseline models. The method's generalizability is further validated through its application to post-training embodied VLA policies.
Methodology
The authors propose World Model RL (WMRL), which replaces real environment execution with a simulated world model to generate execution outcomes. They implement Online Debiasing and Inverse-Variance Denoising to correct biases and reduce noise in the model's predictions. Theoretical analysis is provided to support the convergence guarantees of the proposed methods.
Results
WMRL accelerates training by 3-4 times on various AutoResearch tasks and scales, while achieving performance that exceeds that of larger open-weight models (48B and 120B parameters). The proposed methods also demonstrate generalizability to post-training embodied VLA policies.
Implications
The findings suggest that using world models can significantly enhance the efficiency of training RL agents, making it feasible to scale AutoResearch applications. This approach may also be applicable to other domains requiring efficient RL training, potentially leading to advancements in automated empirical research and other AI-driven tasks.
H-VAEP and H-xT: Valuing Offensive On-the-Ball Actions in Handball by Estimating Probabilities
Theory
- Introduction of H-xT and H-VAEP frameworks tailored for handball analytics.
- Development of a handball-native court zoning layout for improved action valuation.
- Demonstration of the robustness of H-xT compared to traditional rectangular grids.
- H-VAEP yields stable and intuitive player ratings that highlight contributions to build-up play.
Read more
H-VAEP and H-xT: Valuing Offensive On-the-Ball Actions in Handball by Estimating Probabilities
Summary
This paper presents the first comprehensive adaptation of Expected Threat (xT) and Valuing Actions by Estimating Probabilities (VAEP) frameworks for handball, addressing the limitations of traditional player evaluation methods that rely on basic box-score metrics. The authors utilize five seasons of tracking-derived event data from the Handball Bundesliga to develop Handball-xT (H-xT) and Handball-VAEP (H-VAEP). H-xT employs a handball-specific court zoning layout, which is shown to be more robust than standard rectangular grids. H-VAEP is optimized to suit the rapid dynamics of handball while minimizing team-identity leakage. The evaluation demonstrates that H-VAEP provides stable, discriminative, and intuitive player ratings that emphasize the importance of build-up play. The authors also release their code repository to facilitate the practical application of these models in professional handball clubs.
Methodology
The authors adapted the xT and VAEP frameworks to handball by utilizing a handball-specific court zoning layout and optimizing the feature space and context length for H-VAEP. They conducted simulations to validate the robustness of H-xT and evaluated the performance of H-VAEP against traditional metrics using a comprehensive validation framework.
Results
The study found that H-xT is systematically more robust than standard rectangular grids for action valuation in handball. H-VAEP provided player ratings that were stable, reliable, and intuitive, effectively capturing the contributions of players in build-up play. The frameworks were validated against five seasons of event data, demonstrating strong face validity and discrimination compared to traditional performance indicators.
Implications
The development of H-xT and H-VAEP has significant implications for player recruitment, game tactics, and overall performance analysis in handball. By providing a more nuanced understanding of player contributions, these frameworks can help teams make data-driven decisions and improve their competitive edge.
Novel Knowledge-Guided Generative Methods for Synthetic Transcriptomic Data
Generative Models
Graph Learning
- Synthetic data generation can mitigate challenges in accessing high-quality transcriptomic data.
- Integrating biological knowledge through gene graphs enhances the realism and utility of synthetic data.
- MK-TGAN outperforms other generative models in terms of data realism and biological coherence.
- The study highlights the importance of explicit knowledge integration mechanisms in generative models.
Read more
Novel Knowledge-Guided Generative Methods for Synthetic Transcriptomic Data
Summary
This paper addresses the challenges of generating synthetic transcriptomic data, which is crucial for biomedical research but often limited by issues such as data imbalance, biases, and ethical constraints. The authors present a comparative analysis of generative models, particularly focusing on the integration of prior biological knowledge through gene graphs to enhance the realism and utility of synthetic data. They introduce three variants of Generative Adversarial Networks (GANs), with a particular emphasis on MK-TGAN, a multi-kernel, Graph Neural Network-based model that effectively incorporates biological knowledge. The study demonstrates that integrating prior knowledge significantly improves the performance of generative models, with MK-TGAN achieving superior realism and biological plausibility in the generated samples. The evaluation is conducted on breast cancer transcriptomic data, specifically targeting Epithelial-to-Mesenchymal Transition (EMT)-related genes, providing a biologically relevant context for assessing the effectiveness of different knowledge integration strategies.
Methodology
The authors developed three GAN variants: Graph-Modulated GAN (GM-GAN), Graph-Regularized GAN (GR-GAN), and MK-TGAN. GM-GAN and GR-GAN incorporate biological knowledge implicitly, while MK-TGAN explicitly utilizes multiple graph neural network kernels to enhance representation capacity. The models were evaluated using breast cancer transcriptomic data focused on EMT-related genes.
Results
MK-TGAN achieved state-of-the-art performance across metrics such as precision and recall, demonstrating superior fidelity, diversity, and biological coherence in the generated synthetic samples compared to other models.
Implications
The findings suggest that knowledge-guided generative methods can significantly improve the generation of synthetic transcriptomic data, which may facilitate research in genomics and personalized medicine by providing high-quality datasets without ethical constraints.
Fashion Outfit Generation via Unified Sequential Composition Models
Generative Models
Reinforcement Learning
Optimization
- Introduces Constrained Ensemble Generation (CEG) as a formal framework for fashion outfit generation.
- Develops the Unified Sequential Composition Model (USCM) to jointly model compatibility and composition intents.
- Implements Latent Expansion Monte Carlo Tree Search (LE-MCTS) for efficient item retrieval and ensemble synthesis.
- Achieves state-of-the-art results on multiple fashion datasets, demonstrating superior performance in aesthetic and structural evaluations.
Read more
Fashion Outfit Generation via Unified Sequential Composition Models
Summary
This paper addresses the challenge of fashion outfit generation, which involves synthesizing stylistically coherent ensembles from extensive item libraries. The authors formalize this task as Constrained Ensemble Generation (CEG) and model it as a finite-horizon deterministic Markov Decision Process (MDP). They propose the Unified Sequential Composition Model (USCM) that jointly learns set-level compatibility and latent composition intents. To facilitate item retrieval during composition, they introduce the Latent Expansion Monte Carlo Tree Search (LE-MCTS) mechanism, which balances local aesthetic synergy with global structural integrity. The framework is evaluated on the Polyvore Outfits dataset and demonstrates state-of-the-art performance in human preference evaluations, automated aesthetic proxies, and structural validity metrics. The proposed methods effectively address the combinatorial explosion of possible outfits and the non-monotonic nature of fashion compatibility, providing a robust solution for generating high-quality fashion ensembles.
Methodology
The authors model the fashion outfit generation task as a finite-horizon deterministic Markov Decision Process (MDP). The Unified Sequential Composition Model (USCM) is employed to learn compatibility evaluations and policy priors using a dual-pathway Transformer architecture. The Latent Expansion Monte Carlo Tree Search (LE-MCTS) is utilized for item selection, leveraging the learned policy intent and compatibility scores to guide the search process.
Results
The proposed framework achieves state-of-the-art performance on the Polyvore Outfits dataset and demonstrates strong results in zero-shot evaluations on the iFashion and PolyvoreU datasets. The USCM shows superior representational capacity, leading to improved aesthetic compatibility and structural validity in generated outfits.
Implications
This research has significant implications for the fashion retail industry, enabling automated outfit generation that can enhance customer experience and streamline inventory management. The methodologies developed could also be applied to other domains requiring ensemble generation, such as playlist creation or interior design.
Exploring Oversmoothing with Householder Matrices
Graph Learning
- Introduces Householder Graph Neural Network (HouseGNN) to combat oversmoothing in GNNs.
- Uses Householder matrices for node embedding updates, preserving Euclidean norms.
- Proves properties of Householder reflectors that contribute to maintaining node representation diversity.
- Offers a novel perspective on the oversmoothing issue by utilizing orthogonal transformations.
Read more
Exploring Oversmoothing with Householder Matrices
Summary
This paper addresses the issue of oversmoothing in deep graph neural networks (GNNs), where node representations converge towards a low-information subspace as the network depth increases. The author introduces the Householder Graph Neural Network (HouseGNN), which employs Householder matrices to update node embeddings. Instead of directly updating hidden states like traditional GCNs, HouseGNN uses aggregated neighborhood messages to estimate a reflection direction. The node embedding is then updated using a Householder reflector followed by a GroupSort operation, ensuring that the Euclidean norm is preserved at every node and depth. The paper proves three key properties: (i) each internal layer maintains the node-wise Euclidean norm, (ii) the Householder reflector is invariant to scale and sign in the message, and (iii) pairwise distances between nodes can vary due to mismatches in node-wise orthogonal operators. This approach aims to mitigate the oversmoothing phenomenon and enhance the discriminative power of deep GNNs.
Methodology
The methodology involves using Householder matrices to define a reflection direction based on neighborhood messages, followed by updating the node state through a Householder reflection and a GroupSort operation. This approach differs from standard GCNs, which directly update hidden states, and instead focuses on maintaining the orthogonality and norm of node representations.
Results
The proposed HouseGNN demonstrates that by employing Householder reflectors, it effectively mitigates the oversmoothing problem, allowing for better separation of node representations even at greater depths. The theoretical properties established in the paper support the effectiveness of this approach in preserving the discriminative variation across nodes.
Implications
The findings suggest that using Householder transformations in GNNs can lead to more robust models that maintain representation diversity, potentially improving performance in tasks such as node classification and link prediction. This approach may inspire further research into orthogonal transformations in neural networks.
Simulation-to-real transfer learning for infrared spectroscopic chemical sensing and analysis from molecules to complex samples
Efficient ML
- UltraIR is a foundation model for IR spectroscopy with over 100 million parameters.
- The model utilizes simulation-to-real transfer learning to improve chemical inference from IR spectra.
- UltraIR outperforms traditional methods in various chemical sensing tasks and real-world applications.
- The approach allows for effective adaptation using limited labeled experimental spectra.
Read more
Simulation-to-real transfer learning for infrared spectroscopic chemical sensing and analysis from molecules to complex samples
Summary
This paper presents UltraIR, a foundation model for infrared (IR) spectroscopy designed to enhance chemical sensing and analysis through simulation-to-real transfer learning. Traditional methods for interpreting IR spectra are labor-intensive and rely heavily on expert knowledge, making them less effective for unfamiliar compounds and complex samples. UltraIR addresses these limitations by leveraging a large-scale pretraining approach on approximately 60 million simulated IR spectra, employing three complementary objectives: spectral reconstruction, molecular fingerprint similarity alignment, and functional-group prediction. The model is then fine-tuned for specific tasks using limited labeled experimental spectra. The authors demonstrate UltraIR's effectiveness across various benchmarks, including functional-group prediction and mixture-component identification, as well as real-world applications such as bacterial classification and soil property prediction. The results indicate that UltraIR significantly outperforms conventional machine-learning methods and task-specific deep-learning baselines, showcasing its potential for adaptable and data-efficient chemical sensing systems.
Methodology
UltraIR employs a foundation model architecture with pretraining on simulated IR spectra. It uses three objectives for pretraining: spectral reconstruction, molecular fingerprint similarity alignment, and functional-group prediction. The pretrained model is then adapted to specific downstream tasks using limited labeled experimental data.
Results
UltraIR demonstrated superior performance in benchmark evaluations for functional-group prediction, molecular structure elucidation, and mixture-component identification. In real-world applications, it effectively classified bacteria, traced the geographic origin of medicinal herbs, quantified constituents, and predicted soil properties, outperforming conventional and task-specific deep-learning baselines.
Implications
The findings suggest that UltraIR can facilitate scalable and reliable chemical analysis in various fields, including environmental monitoring and clinical analysis, by providing a robust framework for interpreting complex IR spectra with limited labeled data.
Sparse Orthogonal Regression Technique: A Spectral Framework for Equation Discovery, Approximation, and Integration
Theory
- SORT provides a framework for estimating sparse expansions in orthonormal bases from noisy data.
- The technique is particularly effective for discovering ordinary differential equations and supports nonlinear approximation.
- SORT demonstrates improved stability under sparse sampling and noisy conditions compared to traditional methods.
- The method allows for order-consistent model growth, maintaining low-order coefficients as model complexity increases.
Read more
Sparse Orthogonal Regression Technique: A Spectral Framework for Equation Discovery, Approximation, and Integration
Summary
The paper introduces the Sparse Orthogonal Regression Technique (SORT), a novel framework designed for learning orthonormal-basis expansions from noisy and irregularly sampled data. SORT employs L1-regularized regression to estimate expansion coefficients directly from observations, eliminating the need for explicit quadrature or analytic inner-product evaluations. The primary application of SORT is in the data-driven discovery of ordinary differential equations, where vector fields are represented in selected orthogonal bases and learned as sparse coefficient expansions. This approach offers a complementary method to traditional symbolic regression and SINDy-style sparse identification by first establishing a compact spectral representation that can facilitate searches for simpler analytic forms. The authors demonstrate that SORT performs comparably or better than existing library-based sparse-regression methods, particularly when the orthogonal basis is well-suited to the problem. The technique also exhibits greater stability under conditions of sparse sampling, noisy derivative estimates, and representation mismatch. Furthermore, SORT supports nonlinear approximation and the estimation of complex, high-dimensional integrals through coefficient readout, making it a versatile tool for system identification, approximation, and integration. Overall, SORT emphasizes the importance of basis design in scientific modeling, providing a reusable representation that enhances the robustness and interpretability of learned models.
Methodology
SORT combines orthonormal basis representations with sparsity-promoting regression techniques. It estimates coefficients from sampled observations using L1-regularized regression, allowing for the recovery of sparse expansions without the need for explicit quadrature or inner-product evaluations. The framework is evaluated across various tasks, including dynamical-system identification, numerical integration, and nonlinear approximation.
Results
The experiments indicate that SORT matches or surpasses the performance of dense least-squares, kernel-based, and library-based sparse-regression baselines, particularly when the orthogonal basis aligns well with the problem. The results also show that dominant low-order coefficients remain stable as model order increases, supporting the concept of order-consistent model growth.
Implications
SORT has significant implications for scientific computing and machine learning, particularly in areas requiring robust function approximation, system identification, and numerical integration from irregularly sampled data. Its emphasis on basis design can lead to more interpretable and reliable models in various applications.
Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View
Reinforcement Learning
Generative Models
Theory
- Introduces a continuous-time path-space framework for diffusion-RL algorithms.
- Derives a unified policy-gradient estimator that connects disparate RL methods.
- Proposes a multi-sample KDE value-gradient estimator to reduce variance.
- Validates the proposed methods empirically, showing improved performance over existing baselines.
Read more
Designing Reinforcement Learning for Diffusion Models: A Unified Path-Space View
Summary
This paper addresses the challenges of aligning diffusion models with human preferences and task-specific rewards through reinforcement learning (RL) post-training. The authors identify that existing RL algorithms for diffusion models are fragmented, with reverse-trajectory methods relying on discretized likelihood ratios and forward-matching methods using reward-labeled noising versions of rollout samples. The paper introduces a unified path-space framework that reveals the common continuous-time structure underlying these methods. By applying path-space importance sampling to the diffusion-RL objective, the authors derive a policy-gradient estimator that connects stochastic ItΓ΄ integrals with variance-reduced value-gradient forms. This unification allows for a systematic design space organized by value-gradient estimation, weight functions, and sampling choices. The authors propose a multi-sample KDE value-gradient estimator and a scale-bounded weight principle that enhances existing algorithms while maintaining stability. Empirical validation on SD3.5-M and Qwen-Image models demonstrates that the proposed methods effectively close the efficiency gap between Flow-GRPO and DiffusionNFT, outperforming prior diffusion-RL baselines.
Methodology
The authors utilize path-space importance sampling to derive a policy-gradient estimator for the diffusion-RL objective. They analyze the continuous-time structure of existing methods and propose a multi-sample KDE value-gradient estimator, along with a scale-bounded weight principle to guide algorithm design. The framework organizes design choices into a unified template that encompasses various existing algorithms.
Results
The experiments conducted on SD3.5-M and Qwen-Image models demonstrate that the proposed variance-reduction technique effectively closes the efficiency gap between Flow-GRPO and DiffusionNFT. The new weight families proposed track empirical optima, leading to improved performance compared to prior diffusion-RL baselines.
Implications
The unified framework and proposed methods have the potential to streamline the design of reinforcement learning algorithms for diffusion models, making it easier to align generative models with user preferences and task-specific rewards. This could enhance applications in visual generation, including image, video, and 3D synthesis.
A Cloud-Edge System for Multimodal Clinical Screening in Resource-Constrained Rural Settings
Multimodal
Efficient ML
Interpretability
- Introduces a cloud-edge collaborative architecture for multimodal clinical screening.
- Utilizes lightweight edge models for data processing and a cloud LLM for reasoning.
- Achieves high diagnostic accuracy and factual grounding while minimizing data transmission.
- Demonstrates the system's effectiveness under various simulated bandwidth conditions.
Read more
A Cloud-Edge System for Multimodal Clinical Screening in Resource-Constrained Rural Settings
Summary
This paper presents a novel cloud-edge collaborative architecture designed to enhance multimodal clinical screening in resource-constrained rural settings. The authors identify significant barriers to effective medical AI deployment in these areas, including limited bandwidth, computational resources, and the need for integrating diverse medical modalities. To address these challenges, the proposed system utilizes lightweight, domain-specific models at the edge to convert raw medical data into structured outputs, while a cloud-based large language model (LLM) synthesizes these outputs into coherent clinical summaries. An orchestrator on the edge dynamically selects diagnostic tools based on patient context, ensuring relevant modalities are utilized without unnecessary data processing. The evaluation of the system involved 100 multimodal clinical cases across various medical domains, simulating different network conditions. The results demonstrate that the hybrid architecture achieves high oracle accuracy and factual grounding while significantly reducing data transmission size compared to cloud-only systems. The findings emphasize the importance of architectural design in improving clinical decision-making and efficiency in rural healthcare settings.
Methodology
The study employs a cloud-edge collaborative system where edge models process medical data into structured outputs, and a cloud-based LLM synthesizes these outputs into clinical summaries. The system was evaluated using 100 clinical cases under three simulated network profiles, focusing on bandwidth constraints and clinical quality metrics.
Results
The hybrid system achieved oracle accuracy ranging from 0.87 to 0.90, with knowledge graph precision up to 0.96 and coverage precision between 0.95 and 0.99. It maintained a bandwidth-invariant latency of 25-38 seconds while transmitting approximately 6.5 KB of structured evidence, significantly less than cloud-only baselines.
Implications
This research has significant implications for improving healthcare delivery in rural areas by enabling efficient and accurate clinical decision-making through advanced AI systems. The proposed architecture could facilitate better access to specialist-level diagnostics and reduce patient attrition due to follow-up visits.
Smart routes: a system for development and comparison of algorithms for solving vehicle routing problems with realistic constraints
Optimization
- Introduction of the Smart Routes platform for route optimization.
- Comparison of exact and heuristic algorithms for CVRPTW.
- Heuristic and deep learning methods show competitive performance against exact solvers.
- Significant reduction in computation time with heuristic approaches for larger problem sizes.
Read more
Smart routes: a system for development and comparison of algorithms for solving vehicle routing problems with realistic constraints
Summary
This paper addresses the increasingly relevant issue of route optimization in the context of the Capacitated Vehicle Routing Problem with Time Windows (CVRPTW), particularly as urban populations grow. The authors highlight the challenges of applying exact algorithms due to their exponential complexity as problem sizes increase. They present the Smart Routes platform, which facilitates the development and comparison of various algorithms, including exact solvers like SCIP and heuristic methods such as LKH, 2-OPT, 3-OPT, and the deep learning model JAMPR. The study demonstrates that for problem sizes of 50, the performance of deep learning and classical heuristics approaches closely matches that of SCIP while requiring significantly less computation time. For larger problems (size 100), SCIP is approximately 13 times slower than the heuristic methods, which yield similar route costs but with a 50% worse first feasible solution time. The paper emphasizes the need for a unified platform that allows for the integration of custom algorithms and datasets, thus providing a comprehensive tool for both experienced and novice users in transport logistics. The findings suggest that the Smart Routes system can effectively optimize route problems while accommodating various constraints, making it a valuable resource for logistics optimization.
Methodology
The authors developed the Smart Routes platform to implement and compare various algorithms for solving the CVRPTW. They utilized exact solvers (SCIP), heuristic algorithms (LKH, 2-OPT, 3-OPT), and a deep learning model (JAMPR) to evaluate their performance across different problem sizes and constraints. The effectiveness of these methods was assessed based on solution optimization time and final route cost.
Results
The study found that for problem sizes of 50, heuristic and deep learning methods achieved results comparable to the exact solver SCIP but with significantly reduced computation times. For larger problems (size 100), SCIP was found to be approximately 13 times slower than the heuristic methods, which provided similar route costs but with a worse first feasible solution time.
Implications
The Smart Routes platform has the potential to enhance the efficiency of logistics operations by providing a comprehensive tool for route optimization that accommodates various constraints. This can lead to improved delivery times and resource utilization in urban transport logistics.
A Compositional Theory of Curvature in Probabilistic Circuits
Generative Models
Optimization
Theory
- Probabilistic Circuits (PCs) allow for exact inference and tractable curvature measures unlike deep neural networks.
- Global sharpness regularization can lead to underfitting in PCs due to the compositional nature of their curvature.
- The contribution of each sum node to the Hessian trace can be decomposed into contextual usage and local curvature.
- An adaptive sharpness-aware regularizer improves generalization by focusing on nodes with high intrinsic local curvature.
Read more
A Compositional Theory of Curvature in Probabilistic Circuits
Summary
This paper presents a novel perspective on the curvature of loss surfaces in Probabilistic Circuits (PCs), which are generative models that allow for exact inference. The authors argue that treating curvature as a global regularizer can be misleading due to the compositional nature of PCs. They demonstrate that the contribution of each sum node to the Hessian trace can be decomposed into two components: the circuit flow, which indicates how heavily a node is utilized, and a local sharpness term that reflects the node's output distribution. This decomposition reveals that sharpness is contextually influenced and suggests that global regularization may lead to underfitting. To address this, the authors propose an adaptive sharpness-aware regularizer that penalizes nodes based on their intrinsic local curvature while maintaining efficient updates. Empirical results show that this targeted approach improves generalization compared to traditional global regularization methods, particularly in low-data scenarios.
Methodology
The authors analyze the curvature of loss surfaces in PCs by decomposing the contribution of sum nodes to the Hessian trace into two distinct components: circuit flow and local sharpness. They develop an adaptive regularization technique that targets nodes based on their local curvature, allowing for efficient updates while preserving the tractability of the learning process.
Results
The proposed adaptive sharpness-aware regularizer demonstrated improved generalization performance in empirical tests, particularly in low-data regimes, compared to traditional global regularization methods. The results indicated that while global regularization can achieve flatter optima, it may sacrifice training and test log-likelihood, leading to underfitting.
Implications
This work suggests that more nuanced approaches to regularization in generative models like PCs can enhance their performance, particularly in scenarios with limited data. The findings may influence future research on model training strategies and the design of regularization techniques in machine learning.
Decentralized Multi-Player Q-Learning in Episodic Markov Decision Processes with Information Asymmetry
Reinforcement Learning
Theory
Robotics
- Introduces decentralized multi-player Q-learning under information asymmetry.
- Proposes algorithms mQ-learning and mQ-learning-intervals for common rewards and independent rewards scenarios.
- Demonstrates that decentralized learning can achieve regret bounds similar to centralized benchmarks.
- Highlights the challenges posed by exponential growth in joint action space with increasing players.
Read more
Decentralized Multi-Player Q-Learning in Episodic Markov Decision Processes with Information Asymmetry
Summary
This paper investigates decentralized multi-player reinforcement learning in episodic tabular Markov Decision Processes (MDPs) under conditions of information asymmetry. The authors categorize the problem into three scenarios: (A) unobserved actions with common rewards, (B) observed actions with independent rewards, and (C) unobserved actions with independent rewards. They propose two algorithms, mQ-learning and mQ-learning-intervals, for Problems A and B, achieving a regret bound of ΛO(βH4SAjoint T). For Problem C, they introduce mEXC and mEXC-Bellman, which are explore-then-commit algorithms with a regret bound of ΛO(H(SAjoint)1/3T 2/3). The results indicate that decentralized learning under information asymmetry can match the performance of centralized joint-action Q-learning, with regret bounds that are polynomially close to single-agent Q-learning rates. The findings highlight the potential for decentralized learning frameworks in cooperative systems where players cannot communicate during learning but can agree on a protocol in advance.
Methodology
The authors develop decentralized algorithms tailored for three types of information asymmetry in multi-player settings. They utilize a lexicographic ordering of joint actions and maintain confidence bounds to facilitate coordination among players without communication. The algorithms are designed to explore and commit to strategies based on observed rewards and actions, ensuring efficient learning in episodic MDPs.
Results
The proposed algorithms achieve regret bounds of ΛO(βH4SAjoint T) for Problems A and B, and ΛO(H(SAjoint)1/3T 2/3) for Problem C. These results show that decentralized learning can perform comparably to centralized approaches, with regret rates that are logarithmically close to those of single-agent Q-learning, despite the challenges introduced by information asymmetry.
Implications
The findings suggest that decentralized reinforcement learning can be effectively applied in scenarios where agents must operate independently without communication, such as in multi-robot coordination and distributed resource allocation. The results also indicate that information asymmetry does not impose significant additional costs on learning efficiency compared to centralized methods.
A Contract-Grade Verifier for LLM-Generated GPU Kernels, and a Native Blackwell Backward for the Gated-Linear-Recurrence Family
Large Language Models
Theory
Optimization
- Introduction of a twelve-gate contract-grade verifier for GPU kernels.
- Audit of 2,638 machine-generated kernels revealing significant correctness issues.
- Development of the first native Blackwell tcgen05 training backward for the GDN family.
- Demonstration of the verifier's effectiveness through independent validation methods.
Read more
A Contract-Grade Verifier for LLM-Generated GPU Kernels, and a Native Blackwell Backward for the Gated-Linear-Recurrence Family
Summary
This paper addresses the reliability of GPU kernels generated by large language models (LLMs), which often report high correctness rates based on loose testing methods. The authors introduce a contract-grade verifier that evaluates GPU kernels against twelve adversarial gates, ensuring rigorous correctness checks beyond simple output comparisons. The verifier was applied to audit 2,638 machine-generated kernels, revealing that 39.5% were fundamentally broken and 62.1% had at least one violation. This starkly contrasts with the results of existing testing methods, which accepted many of these kernels as correct. The authors also present a novel native Blackwell tcgen05 training backward for the gated-linear-recurrence family, verified against a double-precision oracle. The findings suggest that the perceived correctness of generated kernels is often overstated, and implementing tolerance-free contracts could significantly enhance the reliability of kernel generation.
Methodology
The authors developed a contract-grade verifier that operationalizes the Kernel Contracts taxonomy, consisting of twelve adversarial gates. The verifier was used to audit a large corpus of machine-generated kernels, assessing their correctness against a high-precision reference. The methodology included both tolerance-free checks and those with derived tolerances, ensuring a robust evaluation framework.
Results
The verifier found that 39.5% of the audited kernels were broken beyond any tolerance argument, while 62.1% exhibited at least one violation. The standard testing method accepted 1,487 kernels that the verifier rejected, indicating a significant gap in correctness assessment. The authors also verified their own kernel against a double-precision oracle, confirming its correctness.
Implications
The findings suggest that current methods for assessing the correctness of LLM-generated GPU kernels are inadequate, potentially leading to the acceptance of faulty kernels. The introduction of a rigorous verification framework could improve the reliability of GPU kernel generation, impacting fields that rely on high-performance computing and machine learning applications.
SAGE: Surrogate-gradient Adaptation via Attention-Guided Entropy for Spiking Transformers
Efficient ML
Computer Vision
Theory
- Introduction of SAGE, the first uncertainty-aware adaptive surrogate-gradient framework for spiking transformers.
- Utilization of attention entropy to estimate uncertainty and modulate surrogate-gradient slopes during training.
- Demonstrated consistent accuracy improvements over fixed and learnable surrogate baselines.
- Minimal additional training overhead of only 0.03 ms per mini-batch.
Read more
SAGE: Surrogate-gradient Adaptation via Attention-Guided Entropy for Spiking Transformers
Summary
The paper introduces SAGE, a novel surrogate-gradient adaptation mechanism designed for Spiking Neural Networks (SNNs) that utilize Transformer architectures. SNNs are recognized for their energy efficiency due to their event-driven computation, but training them poses challenges due to the non-differentiable nature of the spike function. Traditional surrogate gradients often apply a fixed shape across layers, which can be suboptimal. SAGE addresses this by estimating block-level uncertainty from normalized self-attention entropy and adapting the surrogate-gradient slope during training. This method allows for improved optimization flexibility while maintaining the original model architecture during inference. The authors validate SAGE through experiments on CIFAR-10 and CIFAR-100 datasets, demonstrating that it achieves 1-2% higher accuracy compared to fixed-surrogate baselines, with minimal additional training overhead. The results underscore the effectiveness of using attention-derived uncertainty as a training signal for adaptive surrogate-gradient learning in spiking transformers.
Methodology
SAGE employs an uncertainty-modulated surrogate-gradient mechanism that estimates uncertainty from the dispersion of self-attention entropy across transformer blocks. This allows for adaptive modulation of the surrogate-gradient slope during backpropagation, enhancing the training process without altering the inference model.
Results
Experiments on CIFAR-10 and CIFAR-100 show that SAGE improves accuracy by 1-2% compared to fixed-surrogate baselines, while only adding 0.03 ms of training overhead per mini-batch. The method effectively leverages attention-derived uncertainty to enhance training effectiveness.
Implications
The findings suggest that adaptive surrogate-gradient mechanisms can significantly improve the training of spiking transformers, potentially leading to more efficient and effective models in various applications, particularly in energy-constrained environments.