AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
24
Papers today
8h
Update frequency
7
Days of history
Guide, Not Bind: Why Defeasible Priors Fail in Augmented Lagrangian Causal Discovery
Graph Learning
Optimization
Theory
- Defeasible priors in causal discovery can lead to significant edge suppression due to the early suppression trap.
- The DADU relaxation rule fails to meet necessary conditions for effective adaptive relaxation.
- Correlation-matching objectives can obscure true causal relationships by tying true edges and their reverses to identical costs.
- A new relaxation operator combined with covariance matching improves edge recovery rates significantly.
Read more
Guide, Not Bind: Why Defeasible Priors Fail in Augmented Lagrangian Causal Discovery
Summary
This paper critiques the use of defeasible priors in differentiable causal discovery methods that employ Augmented Lagrangian Methods (ALM) to enforce forbidden-edge constraints. The authors argue that the common assumption that data can adaptively relax these constraints fails due to two main reasons. First, the sequential penalty-ramping mechanism of ALM often suppresses true edges before any counterfactual checks can detect them, leading to a phenomenon termed the 'early suppression trap.' The authors introduce a new relaxation rule, DADU, which they demonstrate violates necessary conditions for effective adaptive relaxation. Second, they show that the standard correlation-matching objective ties true edges and their reverses to identical costs, which is not due to model unidentifiability but rather the loss of variance information in normalized correlation. They propose a fix combining a new relaxation operator and covariance matching, which significantly improves edge recovery rates compared to DADU. However, even with this fix, the majority of the edge's weight is still misallocated to its reverse direction. The authors conclude that while a provably-identifying objective is not sufficient to prevent suppression, it is also not necessary, indicating deeper issues with penalty schedules in causal discovery.
Methodology
The authors conducted extensive experiments involving 3,072 training runs across various graph sizes, analyzing the performance of the DADU relaxation rule and comparing it to a newly proposed relaxation operator combined with covariance matching. They also examined the effects of different objective functions on edge recovery rates.
Results
The proposed relaxation operator outperformed DADU, recovering wrongly-forbidden edges 27 to 52 times more often across 768 paired trials. However, a significant portion of the edge's weight was still misallocated to its reverse direction. The suppression mechanism was found to worsen under a likelihood objective, indicating that the issues are not solely tied to correlation matching.
Implications
The findings have significant implications for the design of causal discovery algorithms, suggesting that practitioners should be cautious in implementing defeasible priors and reconsider the use of penalty schedules in their models. The proposed methods could enhance the accuracy of causal inference in various applications, including social sciences and biomedical research.
Modern Transformers Are Implicit Hybrids: From Functional Differentiation to Principled Hybrid Architecture Design
NLP
Large Language Models
- Introduces RFIS and RPD metrics for analyzing attention head functions in Transformers.
- Establishes a two-type taxonomy of retrieval and positional heads based on frequency contributions.
- Identifies the Global Positional Band (GPBand) as a critical boundary for functional separation.
- Proposes design principles for hybrid architectures that enhance model performance.
Read more
Modern Transformers Are Implicit Hybrids: From Functional Differentiation to Principled Hybrid Architecture Design
Summary
This paper investigates the hybrid architectures of modern Transformers, particularly those combining Full Attention (FA) and Linear Attention (LA). The authors propose a principled approach to hybrid architecture design based on the functional differentiation observed in RoPE-based Transformers. They introduce two metrics: RoPE Frequency Importance Score (RFIS) and RoPE Positional Dependence (RPD), to analyze attention heads and establish a taxonomy of retrieval and positional heads. The study identifies a significant boundary, termed the Global Positional Band (GPBand), which separates these two functions based on frequency contributions. The authors suggest that positional modeling should be localized while global retrieval should be position-independent. They instantiate these principles in a new architecture called Head-wise Hybrid Architecture (HwH), which demonstrates improved performance in language modeling and commonsense reasoning, particularly in zero-shot long-context extrapolation tasks. The findings emphasize the importance of principled hybrid architecture design in advancing Transformer models.
Methodology
The authors employed behavioral probes to analyze head-level functional differentiation in RoPE-based Transformers. They introduced RFIS and RPD metrics to quantify the impact of frequency contributions and positional modulation on attention distributions. The study utilized controlled Transformer models to validate the proposed taxonomy and design principles, culminating in the development of the HwH architecture.
Results
The application of RFIS and RPD revealed a clear separation of retrieval and positional heads, confirming the existence of the GPBand. The HwH architecture, which implements the proposed design principles, achieved superior performance in language modeling and commonsense reasoning tasks, particularly in zero-shot long-context extrapolation, outperforming traditional Transformer models and existing hybrid architectures.
Implications
The findings suggest that a principled approach to hybrid architecture design can significantly enhance the capabilities of Transformer models, particularly in handling long-context tasks. This research could inform future developments in foundation models and improve their efficiency and effectiveness in various NLP applications.
LeanGRPO: Eliminating Redundant Recomputation in Diffusion RL
Reinforcement Learning
Generative Models
Efficient ML
- Identifies update-stage recomputation as a major bottleneck in trajectory-logprob diffusion RL.
- Introduces LeanGRPO, a framework that eliminates redundant recomputation in diffusion RL.
- Presents two complementary training schedules: LeanGRPO-Retain and LeanGRPO-Reweight.
- Achieves up to 1.83× speedup in training without compromising optimization objectives.
Read more
LeanGRPO: Eliminating Redundant Recomputation in Diffusion RL
Summary
The paper introduces LeanGRPO, a novel framework designed to eliminate redundant recomputation in diffusion reinforcement learning (RL) methods, which has become increasingly important in generative models for images and videos. Traditional diffusion RL methods, such as DanceGRPO and FlowGRPO, often recompute selected timesteps during the update phase, leading to unnecessary computational overhead. LeanGRPO addresses this issue by restructuring the data-parallel layout and proposing two recompute-free training schedules: LeanGRPO-Retain and LeanGRPO-Reweight. LeanGRPO-Retain allows for gradient tracking during rollout, reusing computation graphs and saved activations without recomputation, while LeanGRPO-Reweight utilizes provisional gradients and delays synchronization to manage memory usage effectively. The framework is compatible with existing diffusion RL algorithms and demonstrates significant performance improvements, achieving up to 1.83× speedup while maintaining the original optimization objectives. The paper provides extensive experimental validation across various model scales and workflows, showcasing the practical benefits of LeanGRPO in reducing computational costs and memory overhead.
Methodology
The authors restructured the data-parallel layout used in conventional GRPO methods to allow all GPUs to process the same prompt while generating different samples. They introduced two training schedules: LeanGRPO-Retain, which reuses computation graphs during the update phase, and LeanGRPO-Reweight, which uses provisional gradients to manage memory consumption efficiently.
Results
LeanGRPO demonstrated up to 1.83× end-to-end speedup compared to traditional methods while preserving the original optimization objectives. The framework was tested across various model scales and RL workflows, showing significant improvements in computational efficiency and memory management.
Implications
LeanGRPO's approach can significantly enhance the efficiency of diffusion RL methods, making them more practical for real-world applications in generative modeling. The reduction in computational overhead can lead to faster training times and lower resource consumption, which is crucial for scaling up generative models.
Resolution-Aware Experimental Design under Partial Identifiability
Theory
Optimization
- Introduction of Resolution-Aware Experimental Design (RAED) to address partial identifiability.
- Establishment of cross-nuisance structural aliasing as a key challenge in experimental design.
- Development of a learned implementation for RAED with finite-sample calibration.
- Empirical results demonstrate significant differences in experiment selection and structural resolution.
Read more
Resolution-Aware Experimental Design under Partial Identifiability
Summary
This paper introduces Resolution-Aware Experimental Design (RAED), a novel framework for selecting experiments that account for partial identifiability due to nuisance variables. Traditional experimental design focuses on maximizing information gain, but under conditions of persistent nuisance uncertainty, the same observation can yield different structural interpretations. RAED aims to minimize the expected size of the nonempty structural candidate set while controlling the risk of false exclusions. The authors establish a theoretical foundation for RAED, demonstrating that an experiment can be informative yet fail to provide precise structural resolutions due to cross-nuisance structural aliasing. They develop a learned score-based implementation that incorporates finite-sample calibration for nuisance averaging and tail sensitivity. The paper presents empirical evaluations using two subsurface-flow benchmarks, revealing significant differences in experiment selection and structural resolution between RAED and expected information gain approaches. The findings highlight the importance of considering nuisance variables in experimental design to achieve valid structural conclusions.
Methodology
The methodology involves defining a resolution-aware design objective that minimizes the expected size of the structural candidate set while controlling for false exclusions. The authors develop a learned score-based implementation that calibrates for nuisance averaging and tail-sensitive validity, allowing for practical experiment selection based on theoretical principles.
Results
The empirical evaluations showed that RAED and expected information gain led to different experiment selections, with RAED providing better structural resolution in the presence of nuisance variables. In particular, the fluvial benchmark demonstrated that tail protection altered the selected experiments, highlighting the significance of ambiguity in structural conclusions.
Implications
The findings suggest that RAED can enhance experimental design in fields where nuisance variables are prevalent, such as environmental science and biomedical research. By addressing partial identifiability, RAED may lead to more reliable conclusions and better-informed decision-making in complex systems.
Latent Energy Action Planning with World Models
Robotics
Reinforcement Learning
Optimization
- LEAP optimizes complete action horizons through a frozen latent world model.
- The method couples latent-goal matching with decoder-predicted terminal-state matching.
- LEAP achieves a 17.3 percentage-point improvement in mean success over traditional methods.
- The approach retains the efficiency of the LeWM representation while enhancing action selection.
Read more
Latent Energy Action Planning with World Models
Summary
This paper introduces Latent Energy Action Planning (LEAP), a novel approach to model predictive control that enhances the performance of latent world models (LeWM) by optimizing the entire action horizon as a differentiable variable. The authors identify a limitation in existing methods, where optimizing a single learned latent objective can lead to action sequences that do not align with the intended goals. LEAP addresses this by coupling terminal latent goal matching with a terminal-window state energy, ensuring that both the predicted terminal latent and the decoder-predicted terminal descriptor align with the goal descriptor. The method employs a frozen LeWM for action refinement, utilizing a quasi-Newton solver and post-optimization projection to maintain action feasibility. The evaluation across four control domains demonstrates significant improvements in success rates, showcasing LEAP's ability to enhance planning efficiency without retraining the world model.
Methodology
LEAP treats the entire action horizon as a differentiable variable, optimizing it through a frozen LeWM. It utilizes a goal-conditioned proposal for initialization, a quasi-Newton solver for action refinement, and post-optimization projection to ensure admissible actions. The method evaluates both latent and decoder-predicted representations to ensure alignment with the goal.
Results
LEAP achieved a mean success rate of 94.8% across four control domains, compared to 77.5% for the baseline method (LeWM+CEM), representing a 17.3 percentage-point improvement. The planning time added was only 0.08 seconds per trial, indicating high efficiency.
Implications
The findings suggest that LEAP can significantly enhance the performance of robotic control systems and other applications requiring model predictive control by ensuring that planned actions are aligned with desired outcomes. This could lead to more reliable and efficient robotic systems in various domains.
Gradients Know What Outcomes Don't: Unlocking Reinforcement Learning for LLM Reasoning with Gradient-Aligned Rewards
NLP
Large Language Models
Reinforcement Learning
- Introduction of Gradient-Aligned Reward (GAR) for LLM reasoning.
- GAR utilizes cosine similarity in gradient space to provide dense rewards.
- Empirical validation shows GAR improves performance on math benchmarks.
- GAR operates with less than 9% overhead compared to traditional methods.
Read more
Gradients Know What Outcomes Don't: Unlocking Reinforcement Learning for LLM Reasoning with Gradient-Aligned Rewards
Summary
This paper introduces Gradient-Aligned Reward (GAR), a novel reinforcement learning mechanism designed to enhance reasoning capabilities in large language models (LLMs) by addressing the limitations of traditional binary outcome rewards. The authors argue that existing methods fail to differentiate among correct trajectories, leading to ineffective credit assignment. GAR leverages the gradient space of the model by using truncated backpropagation to compute a compact gradient vector for each rollout, which is then compared to an expert-anchor gradient derived from existing chain-of-thought solutions. This cosine similarity provides a dense, reasoning-aware reward that significantly improves the model's ability to discern high-quality reasoning paths. The authors validate GAR through empirical tests on Qwen3-4B and Qwen3-8B models, demonstrating consistent performance improvements on competitive math benchmarks and successful transfer to other tasks without requiring domain-specific data. Overall, GAR represents a significant advancement in the application of reinforcement learning to LLMs, enabling more nuanced and effective reasoning processes.
Methodology
The GAR mechanism employs truncated backpropagation through the output projection layer of the LLM to extract gradient vectors for each rollout. These vectors are then compared to a reference gradient (expert-anchor) using cosine similarity, yielding a reward signal that reflects the reasoning quality of the trajectories. This approach allows for real-time, online computation of rewards without the need for extensive offline annotations or additional models.
Results
The implementation of GAR on Qwen3-4B and Qwen3-8B models resulted in consistent improvements in performance across four competition-level math benchmarks. The method also demonstrated effective transferability to tasks such as GPQA Diamond and MMLU-Pro, achieving these results without the need for domain-specific training data.
Implications
GAR has the potential to significantly enhance the reasoning capabilities of LLMs in various applications, particularly in areas requiring complex problem-solving and logical reasoning. By providing a more nuanced reward mechanism, GAR could lead to advancements in AI systems that require sophisticated reasoning, such as automated tutoring systems, advanced conversational agents, and decision-making tools.
LongCounsel-8: A Benchmark Suite for Longitudinal Depression Tracking from Multi-Session Counseling Dialogues
NLP
- LONGCOUNSEL-8 is a novel benchmark suite for longitudinal depression tracking from multi-session dialogues.
- The benchmark includes 7,749 five-session counseling trajectories with standardized PHQ-8 states.
- Validation tests confirm the fidelity of the constructed states and the integrity of the benchmark.
- Existing methods show lower reliability in tracking worsening depression trends.
Read more
LongCounsel-8: A Benchmark Suite for Longitudinal Depression Tracking from Multi-Session Counseling Dialogues
Summary
The paper introduces LONGCOUNSEL-8, a benchmark suite designed to facilitate longitudinal depression tracking from multi-session counseling dialogues. The authors identify a significant gap in existing resources, which either lack standardized session-level depression labels or provide insufficient longitudinal data. To construct LONGCOUNSEL-8, the authors address three main challenges: ensuring longitudinal consistency and diversity, grounding symptom progression in empirical patterns, and expressing controlled depression states naturally without revealing target labels. The benchmark consists of three independently generated datasets totaling 7,749 five-session counseling trajectories, based on real-world client profiles and depression trajectories. The authors employ techniques such as profile-grounded simulation, empirically informed state construction, and indirect behavioral realization to create the datasets. Validation tests confirm the fidelity of controlled states and the plausibility of counseling language. Experiments reveal that existing depression tracking methods struggle with accurately identifying trends in worsening trajectories, and that additional session history may not always enhance prediction accuracy. Overall, LONGCOUNSEL-8 serves as a foundational resource for advancing the assessment of mental health changes over time.
Methodology
The authors developed LONGCOUNSEL-8 using a combination of profile-grounded longitudinal simulation, empirically informed state construction, and indirect behavioral realization. They created datasets based on real client profiles and depression trajectories, ensuring that the dialogues reflect natural conversational patterns while maintaining controlled depression states.
Results
The validation of LONGCOUNSEL-8 demonstrated that simulated self-reports effectively recover the intended depression states. Experiments on existing depression tracking methods revealed that lower single-session score errors do not guarantee accurate trend identification, particularly for worsening trajectories. Additionally, incorporating more session history did not consistently improve prediction accuracy.
Implications
LONGCOUNSEL-8 has the potential to enhance the development and evaluation of machine learning models for longitudinal mental health assessment, moving beyond static predictions to more reliable tracking of depression changes over time. This benchmark can support researchers in creating more effective interventions and tools for mental health professionals.
A Large Open Multi-Energy Corpus of Soil Compaction Tests, with Machine-Learning Baselines
Theory
Optimization
- Introduction of a large, open dataset of soil compaction tests from multiple sources.
- Identification of physical impossibilities in existing compaction data.
- Establishment of a baseline for optimum degree of saturation in soil compaction.
- Application of machine learning models to predict compaction parameters with varying degrees of accuracy.
Read more
A Large Open Multi-Energy Corpus of Soil Compaction Tests, with Machine-Learning Baselines
Summary
This paper presents a comprehensive dataset of soil compaction tests, addressing the limitations of existing datasets that are often restricted to a small number of specimens from single laboratories and compactive energy levels. The authors compiled a corpus containing 2,854 laboratory compaction tests sourced from six public datasets, covering 162 provenance groups and four Proctor energy levels. Each record adheres to the Proctor method, ensuring reliability and consistency. The study highlights the physical impossibility of certain published compaction data and establishes a baseline for the optimum degree of saturation across the corpus. The authors employed machine learning techniques, including tabular foundation models and symbolic regression, to predict maximum dry density (MDD) and optimum moisture content (OMC). The results indicate that while compactive energy has a negligible marginal effect, it is conditionally significant. The study also emphasizes the importance of screening predictions against the zero-air-voids line to ensure physical admissibility. Overall, this research provides a valuable resource for future studies in soil compaction and machine learning applications in geotechnical engineering.
Methodology
The authors compiled a dataset of soil compaction tests from various public sources, ensuring adherence to the Proctor method. They applied machine learning techniques, including tabular foundation models and symbolic regression, to estimate maximum dry density and optimum moisture content based on the compiled data. The models were evaluated using different validation strategies, including random folds and provenance-based folds.
Results
The tabular foundation model achieved R² values of 0.824 for density and 0.784 for water content under random folds. When folds were drawn around provenance, the R² values decreased to 0.727 and 0.696, respectively. The model's performance dropped further to 0.520 and 0.614 when a whole source was held out. The inclusion of compactive energy improved the prediction accuracy for modified-Proctor records, achieving an R² of 0.740.
Implications
The dataset and findings can significantly enhance the understanding of soil compaction characteristics and improve predictive modeling in geotechnical engineering. The research may lead to more efficient compaction control practices and better-informed engineering decisions regarding soil fill materials.
Constant regret in general games via higher-order optimism
Theory
Optimization
- Introduction of the HOOD algorithm, achieving O(N^3 log^2 K) individual regret.
- The algorithm combines higher-order optimism with entropic regularization to control oscillations in play.
- HOOD guarantees constant regret for all players in arbitrary N-player games.
- The algorithm is horizon-free, meaning it does not require prior knowledge of the play duration.
Read more
Constant regret in general games via higher-order optimism
Summary
This paper introduces a novel uncoupled learning algorithm named Higher-Order Optimism with Discounting (HOOD) that guarantees O(N^3 log^2 K) individual regret for each player in an N-player normal form game with up to K actions per player. The algorithm is a variant of the optimistic follow-the-regularized-leader (OptFTRL) method, incorporating a discounted (N + 1)-th order predictor and entropic regularization to mitigate large oscillations in the sequence of play. This approach addresses a significant challenge in achieving constant regret in general games, where previous methods have struggled. The authors demonstrate that if all players adopt the HOOD algorithm, their individual regret remains bounded over time, leading to convergence of the empirical distribution of play to the set of coarse correlated equilibria at a rate of O(1/T). The algorithm operates in a deterministic, full-information setting, where players receive feedback in the form of their mixed payoff vector, allowing for effective strategy updates without dependence on other players' payoff functions.
Methodology
The authors developed the HOOD algorithm, which utilizes a discounted (N + 1)-th order predictor and entropic regularization to dampen oscillations in player strategies. The algorithm is designed to be uncoupled, meaning each player's strategy updates do not depend on the payoff functions of other players, but rather on their own feedback.
Results
The main result shows that if all players in a finite N-player game with up to K actions follow the HOOD algorithm, their individual regret is bounded by O(N^3 log^2 K) for all time horizons. This result is significant as it confirms the possibility of achieving constant individual regret in general games, a long-standing question in the field.
Implications
The findings suggest that the HOOD algorithm can be applied to various multi-agent systems and game-theoretic scenarios, providing a robust framework for players to learn and adapt their strategies over time without accumulating regret. This could enhance the design of algorithms in competitive environments, such as online platforms, auctions, and resource allocation problems.
Free Pause Tokens
NLP
Large Language Models
Efficient ML
- Introduces the 'free pause token' concept for efficient state-prediction separation in language models.
- Reduces computational costs associated with SPS from 1.9x to as low as 1.09x pretraining FLOPs.
- Implements four mechanisms to enhance training efficiency and maintain model performance.
- Achieves significant improvements in next-token prediction accuracy with minimal added inference costs.
Read more
Free Pause Tokens
Summary
The paper introduces a novel approach to state-prediction separation (SPS) in language models, which traditionally requires significant computational resources. The authors propose the concept of a 'free pause token,' which allows the prediction stream to operate without writing additional keys or values, effectively utilizing existing sequence positions. This innovation significantly reduces the computational overhead associated with SPS, bringing the cost down from approximately 1.9 times the pretraining FLOPs to as low as 1.09 times, depending on the trade-off between quality and compute. The method leverages four key mechanisms to enhance efficiency: a two-pass split for FlashAttention compatibility, a zero-length prediction window, a shared gated feedforward network, and phasing the separation to the end of the training run. The results demonstrate that this approach improves next-token prediction accuracy while maintaining low inference costs, making it a promising advancement in the efficiency of large language models.
Methodology
The authors utilize a two-stream architecture where the state stream summarizes context and the prediction stream generates next-token predictions without writing new keys or values. This is achieved through a combination of FlashAttention-friendly passes, a zero-length prediction window, a shared gated feedforward network, and strategic phasing of the separation during training.
Results
The proposed method leads to a 2-3 centinat improvement in next-token prediction accuracy on a 1B parameter model. The wall-clock training time overhead is reduced to 1.33x compared to traditional methods, with potential reductions to 1.09x when trading off some performance.
Implications
This work has significant implications for the efficiency of training large language models, potentially enabling faster training times and lower resource consumption while maintaining or improving model performance. It opens avenues for further research into optimizing transformer architectures and could influence future developments in NLP applications.
From Nowcasting to Forecasting: Adapting a Reanalysis-Trained Cloud Cover Model to Observations
Generative Models
Time Series
Computer Vision
- CloudCast v2 improves cloud-cover forecasting accuracy over its predecessor by 10%.
- The model effectively integrates reanalysis data with satellite observations for enhanced predictions.
- It retains spatial detail from satellite cloud fields, addressing a key challenge in operational forecasting.
- The model demonstrates improved performance in longer lead times (1-12 hours) compared to traditional methods.
Read more
From Nowcasting to Forecasting: Adapting a Reanalysis-Trained Cloud Cover Model to Observations
Summary
This paper presents CloudCast v2, a machine-learning model designed for 12-hour cloud-cover forecasting, which adapts a model trained on reanalysis data to satellite-derived observations. Accurate cloud-cover forecasts are crucial for various applications, including temperature prediction and solar-power management. The authors highlight the limitations of existing forecasting methods, which struggle to maintain accuracy as cloud fields evolve over longer lead times. CloudCast v2 is trained on the Copernicus European Regional Reanalysis to learn cloud-evolution dynamics and utilizes conditional flow matching to generate forecasts based on observed cloud fields and numerical weather prediction (NWP) inputs. The model demonstrates a 10% reduction in mean absolute error compared to its predecessor, CloudCast v1, and shows improved spatial agreement in cloud cover predictions after 3-6 hours. This research indicates that observation-initialized machine-learning forecasts can effectively extend beyond the typical nowcasting range while preserving spatial details from satellite data.
Methodology
CloudCast v2 is developed by first training a vision transformer-based model on reanalysis data to learn cloud evolution dynamics. The model is then adapted to satellite-derived cloud fields using conditional flow matching, a generative method that transforms noise into cloud-cover forecasts conditioned on observed initial cloud fields and NWP inputs.
Results
CloudCast v2 achieves a 10% reduction in mean absolute error compared to CloudCast v1 and surpasses it in fractions skill score after approximately 3-6 hours, depending on cloudiness categories. This indicates a significant improvement in both accuracy and spatial agreement in cloud cover forecasts.
Implications
The findings suggest that integrating machine learning with observational data can enhance the accuracy of cloud-cover forecasts, which is vital for meteorological predictions and applications like solar energy management. This approach may also pave the way for more effective forecasting models in other meteorological contexts.
VestigeKV: The NoPE-MLA KV Cache Carries Its Own Eviction Signal in a Vestigial Branch
Large Language Models
Efficient ML
Theory
- VestigeKV introduces a novel method for managing KV caches that utilizes a query-independent eviction signal.
- The architecture allows for efficient partitioning of cache rows, retaining relevant data while archiving others without deletion.
- Retrieval performance remains high (1.00 at 8× and 0.92 at 32× compression) with no changes to model weights or kernels.
- The paper provides theoretical insights into the mathematics of query-independent salience and its implications for cache management.
Read more
VestigeKV: The NoPE-MLA KV Cache Carries Its Own Eviction Signal in a Vestigial Branch
Summary
This paper addresses the challenge of efficiently managing long-lived key-value (KV) caches in machine learning models, particularly in scenarios where queries that will access the cache are not known at the time of cache compression. The proposed solution, VestigeKV, utilizes a query-independent eviction signal derived from a 64-dimensional decoupled branch, which is repurposed from the RoPE (Rotary Positional Encoding) architecture into a salience channel. This method allows for effective cache partitioning, where the most relevant rows are retained in an attended tier while others are archived without deletion. The architecture maintains retrieval performance at 1.00 under 8× compression and 0.92 under 32× compression, demonstrating significant efficiency gains without requiring any changes to model weights or kernels. The paper also discusses the theoretical foundations that enable query-independent salience and the engineering aspects of implementing the cache layout and recall tier. Overall, VestigeKV provides a novel approach to cache management that preserves retrieval accuracy while minimizing computational overhead.
Methodology
The methodology involves leveraging a 64-dimensional decoupled branch from the NoPE (No Positional Encoding) MLA (Multi-Layer Attention) model to create a salience channel for cache eviction. The cache is partitioned based on this signal, allowing for the retention of the most relevant rows while archiving others. The implementation requires reading only 11% of each row during compression events, maintaining the original decoding path unchanged.
Results
The results indicate that VestigeKV achieves retrieval rates of 1.00 under 8× compression and 0.92 under 32× compression, with a negligible footprint increase. The attended tier occupies only 0.25 KB of the total cache size, and the archived rows remain bit-exact and GPU-resident, ensuring efficient access during model execution.
Implications
The implications of this research extend to improving the efficiency of KV cache management in large language models and other machine learning applications, particularly in scenarios where query patterns are unpredictable. The method could enhance the performance of models by ensuring that relevant data is readily accessible without incurring significant computational costs.
OSR: Output Space Redistribution for Adaptive Label Removal in Classification Models
Efficient ML
Theory
- OSR provides a training-free approach to label removal in classification models.
- The method operates directly in the output space, avoiding the need for retraining or parameter modifications.
- It employs a two-step filter for output confidence redistribution, ensuring model utility post-removal.
- OSR is model-agnostic and requires only lightweight label-level statistics.
Read more
OSR: Output Space Redistribution for Adaptive Label Removal in Classification Models
Summary
The paper addresses the challenge of label removal in classification systems, particularly in contexts where categories must be dynamically updated or eliminated due to evolving taxonomies. Existing methods for label removal, which can be categorized into retraining-based and feature-space-adjustment-based approaches, often suffer from limitations such as high computational costs, reliance on original data, and poor scalability. The authors propose a novel method called Output Space Redistribution (OSR), which operates directly in the output space of classification models. This approach allows for adaptive label removal without the need for retraining or modifications to model parameters. OSR employs a two-step filtering process that projects original output confidence vectors to remove the influence of deprecated labels and redistributes the remaining confidence values to maintain model utility. The method is designed to be model-agnostic and requires only existing labels and prior output confidences, thus addressing privacy concerns associated with data-dependent solutions. Extensive experiments demonstrate that OSR achieves competitive performance compared to full retraining, with significant improvements in computational efficiency and privacy preservation across various classification tasks.
Methodology
The OSR method utilizes a two-step filtering process that first projects the original output confidence vectors to eliminate the impact of deprecated labels, followed by redistributing the remaining confidence values among the active classes. This approach allows for effective label removal without retraining the model or accessing the original training data.
Results
The experiments conducted across four datasets reveal that OSR performs competitively against full retraining methods, demonstrating substantial improvements in computational efficiency and privacy preservation. The results indicate that OSR can effectively adapt classification models to changing label conditions while maintaining model performance.
Implications
The OSR method has significant implications for real-world applications where classification models need to adapt to evolving taxonomies, such as e-commerce and financial classification systems. Its efficiency and privacy-preserving characteristics make it a valuable tool for maintaining model relevance without the overhead of retraining.
Mesh-Native Physics-Informed Graph Surrogates for TCAD-in-the-Loop Design Space Exploration
Graph Learning
Optimization
Efficient ML
- Introduction of a physics-informed GAT surrogate for TCAD simulations.
- Surrogate operates directly on the tetrahedral mesh, enhancing transferability across device geometries.
- Combines data loss with finite-volume current-continuity residuals for physics embedding.
- Achieves significant speedup in design evaluations, particularly for large multi-fin arrays.
Read more
Mesh-Native Physics-Informed Graph Surrogates for TCAD-in-the-Loop Design Space Exploration
Summary
This paper introduces a novel physics-informed graph attention network (GAT) surrogate designed for technology computer-aided design (TCAD) simulations, specifically targeting the challenges of multi-objective design space exploration (DSE) in the context of multi-fin tri-gate GaN FinFETs. Traditional TCAD simulations are computationally intensive, particularly for three-dimensional structures, which limits the ability to explore complex design spaces efficiently. The proposed GAT surrogate operates directly on the tetrahedral mesh of the device geometry, predicting essential electrostatic potential and quasi-Fermi levels at each mesh node. By embedding carrier-transport physics into the training objective, the model not only retains fidelity but also exhibits size generalization, allowing it to be applied to larger device arrays without retraining. The integration of a deep ensemble approach provides per-node uncertainty estimates, facilitating an active-learning loop that screens candidate designs rapidly. The surrogate demonstrates significant improvements in throughput compared to full TCAD simulations, achieving a root mean square error (RMSE) of less than one volt per field against ground-truth results. This advancement enables efficient exploration of the Pareto front across various device scales, which was previously infeasible due to the prohibitive costs of direct TCAD evaluations.
Methodology
The methodology involves training a physics-informed graph attention network that directly processes the tetrahedral mesh of the device. The training combines a data loss function with finite-volume current-continuity residuals, ensuring that the physics of carrier transport is embedded in the learning process. The model's architecture allows for size generalization, meaning it can be applied to larger device configurations without needing retraining. An active-learning loop is implemented, utilizing uncertainty estimates from a deep ensemble to prioritize the most informative designs for full TCAD simulation.
Results
The GAT surrogate was benchmarked against Sentaurus Device simulations for multi-fin tri-gate GaN FinFETs, achieving a sub-volt RMSE for the three drift-diffusion fields compared to ground-truth TCAD results. The surrogate's inference time for large multi-fin arrays was under one second per device, significantly enhancing throughput and enabling efficient exploration of the design space.
Implications
The proposed approach has the potential to revolutionize the design process for advanced semiconductor devices by drastically reducing the computational burden associated with TCAD simulations. This could lead to faster iterations in design and optimization, ultimately accelerating the development of high-performance electronic devices.
Frontier LLMs are effective batch optimizers: Assessing reasoning models in continuous and discrete settings
Large Language Models
Optimization
- Frontier LLMs can act as effective batch optimizers in discrete settings, particularly for molecular optimization tasks.
- Performance of LLMs in continuous optimization tasks is competitive but inconsistent compared to classical methods.
- LLMs show significant advantages in semantically rich environments, leveraging their pretraining data effectively.
- The study underscores the need for further exploration of LLM capabilities in optimization contexts.
Read more
Frontier LLMs are effective batch optimizers: Assessing reasoning models in continuous and discrete settings
Summary
This paper investigates the performance of frontier large language models (LLMs) as batch optimizers in both continuous and discrete optimization settings. The authors focus on the Anthropic family of models and evaluate their effectiveness on a range of black-box optimization tasks, including numerical test functions and molecular optimization. The study reveals that while LLMs can serve as competitive zero-shot batch optimizers, their performance is inconsistent compared to classical optimization methods, particularly in continuous settings. However, in semantically rich discrete settings, such as molecular optimization, LLMs demonstrate superior performance and sample efficiency, often outperforming specialized optimization techniques. The findings highlight the potential of LLMs in navigating complex design spaces, especially where traditional optimization methods may struggle due to high evaluation costs.
Methodology
The authors evaluated the performance of frontier LLMs using a set of well-known optimization functions in continuous settings and molecular optimization tasks in discrete settings. They disguised the domains and ranges of continuous functions to minimize memorization exploitation and used SMILES strings for molecular representations. The performance of LLMs was compared against classical optimization methods, particularly Bayesian Optimization with Gaussian processes.
Results
The results indicate that frontier reasoning LLMs are effective but exhibit brittle performance in numerical optimization tasks, often rivaling classical Bayesian Optimization methods. In contrast, for molecular optimization, LLMs frequently surpassed specialist methods, demonstrating high performance and sample efficiency with limited oracle sampling budgets.
Implications
The findings suggest that frontier LLMs could be leveraged for complex optimization tasks in scientific domains, such as drug discovery, where traditional methods may be limited by high evaluation costs. This research opens avenues for further development of LLM-based optimization strategies and highlights the importance of task-specific adaptations.
Verify Before You Distill: Prompt-Level Teacher Gating for On-Policy Distillation
NLP
Large Language Models
Reinforcement Learning
Efficient ML
- TGOPD introduces prompt-level reliability checks for teacher supervision in OPD.
- The method significantly outperforms Vanilla OPD across multiple domains.
- TGOPD enhances GPU utilization on teacher nodes, reducing idle compute resources.
- The approach mitigates the risk of misleading updates from unreliable teacher signals.
Read more
Verify Before You Distill: Prompt-Level Teacher Gating for On-Policy Distillation
Summary
The paper introduces Teacher-Gated On-Policy Distillation (TGOPD), a novel approach to on-policy distillation (OPD) that enhances the reliability of teacher supervision in training language models. Traditional OPD applies teacher feedback uniformly across prompts without verifying the teacher's reliability, which can lead to misleading updates when the teacher is incorrect. TGOPD addresses this by implementing a prompt-level reliability check using a small set of teacher probes to assess the teacher's performance before applying dense supervision. If the teacher is deemed reliable, dense OPD is used; otherwise, a different method is employed. The authors demonstrate that TGOPD significantly improves performance over Vanilla OPD across various tasks, including mathematics and code generation, while also increasing GPU utilization on the teacher side, thus reducing computational waste. The results indicate that TGOPD not only achieves better accuracy but also optimizes resource usage in asynchronous OPD settings.
Methodology
TGOPD utilizes a set of teacher probes to evaluate the reliability of the teacher's responses at the prompt level. It routes prompts to either dense OPD or a different method based on the reliability assessment, thereby ensuring that only reliable teacher signals are used for training.
Results
TGOPD outperformed Vanilla OPD in all six single-domain settings and achieved higher averages across seven benchmarks in multi-domain training. Additionally, it increased teacher-side GPU utilization from 9.8% to 78.9%, demonstrating improved efficiency in resource usage.
Implications
The findings suggest that incorporating reliability checks in teacher-student frameworks can lead to more effective training of language models, potentially influencing future research and applications in model distillation and reinforcement learning.
High-Dimensional Learning Dynamics of Attention-Indexed Models
Theory
Optimization
Large Language Models
- Introduces a high-dimensional dynamical framework for extensive-rank attention models.
- Establishes a finite-dimensional characterization of population loss in attention-indexed models.
- Demonstrates that tied attention induces automatic symmetry breaking, achieving weak recovery in Θ(d² log d) samples.
- Identifies a fast-slow learning mechanism in untied attention, affecting recovery dynamics based on symmetry breaking.
Read more
High-Dimensional Learning Dynamics of Attention-Indexed Models
Summary
This paper investigates the training dynamics of attention-indexed models, particularly in high-dimensional settings where attention matrices have extensive rank. The authors establish a framework that characterizes the population-loss landscape using a finite set of trace order parameters, contrasting it with the infinite hierarchy of matrix moments governing online stochastic gradient descent (SGD). They demonstrate that the parameterization of attention matrices influences learning dynamics significantly. Specifically, tied attention (S = WW⊺) facilitates automatic symmetry breaking, leading to weak recovery in Θ(d² log d) samples, while untied attention (S = UV⊺) exhibits a fast-slow learning mechanism, where the pre-activation means evolve quickly while overlaps evolve slowly. This framework provides insights into how different parameterizations can either overcome or fail to overcome uninformative states during learning, emphasizing the role of architectural implicit bias in attention mechanisms.
Methodology
The authors develop a theoretical framework for analyzing attention-indexed models in high-dimensional settings. They derive the population loss landscape and the dynamics of online SGD, showing that the training dynamics can be approximated by a finite-dimensional system despite the infinite hierarchy of matrix moments. The analysis includes comparisons of different parameterizations of attention matrices (tied vs. untied) and their effects on learning dynamics.
Results
The study reveals that the population loss landscape can be characterized by a finite number of parameters, while the training dynamics involve an infinite hierarchy of moments. Tied attention facilitates symmetry breaking and leads to effective learning, while untied attention's dynamics depend on the interplay between fast and slow processes, affecting the ability to recover informative features.
Implications
The findings suggest that understanding the dynamics of attention mechanisms in high-dimensional settings can inform the design of more effective learning algorithms. The insights into parameterization effects may lead to improved architectures for foundation models in various applications, particularly in NLP and other domains relying on attention mechanisms.
Equation Recast for Canonical Operator Learning Across Parametric PDEs
Theory
Efficient ML
- Introduces 'equation recast' to reformulate parametric operator learning as a single canonical operator.
- Enables zero-shot predictions across new parameter regimes by analytically deriving operator variations.
- Improves integration of sparse and heterogeneous datasets into a common representation.
- Provides convergence diagnostics to identify unreliable predictions in the learning process.
Read more
Equation Recast for Canonical Operator Learning Across Parametric PDEs
Summary
This paper introduces a novel approach called 'equation recast' for learning solution operators across parametric partial differential equations (PDEs). Traditional data-driven models often struggle with generalizing across varying parameters, leading to unreliable predictions outside the training distribution. The authors propose reformulating the learning task to focus on a single canonical operator, which captures parameter-induced variations analytically derived from the governing equations. This method allows for zero-shot predictions across new parameter regimes by transforming the governing equations into a structured effective source, thereby enhancing the model's ability to generalize from sparse and heterogeneous datasets. The authors demonstrate the effectiveness of their approach through high-fidelity simulations in nuclear fusion, where they unify electron-temperature data across multiple device geometries using a jointly trained operator. The framework not only improves data efficiency but also provides a mechanism for monitoring inference reliability through convergence diagnostics, making it a promising route for developing reusable neural PDE solvers.
Methodology
The authors reformulate the learning of parametric operators by analytically deriving variations from the governing equations and representing them as effective sources. This allows the learning process to focus on approximating a single canonical operator, which can be reused across different parameter configurations. The method employs a fixed-point iteration for inference, where convergence indicates reliable predictions.
Results
The equation recast framework successfully unified electron-temperature data from four different tokamak geometries into a single operator, demonstrating improved generalization capabilities and data efficiency. The method also provided a mechanism for diagnosing the reliability of predictions based on convergence behavior during inference.
Implications
This work has significant implications for scientific machine learning applications, particularly in fields requiring reliable surrogate models for PDEs, such as fusion plasma control and real-time digital twins. The ability to generalize across parameter variations with limited data could enhance the development of robust predictive models in various scientific domains.
Headroom-Drift Replay: A Primitive for Principled Replay Control in GRPO
Reinforcement Learning
Large Language Models
Multimodal
- Introduces Headroom-Drift Replay as a principled approach to replay control in GRPO.
- Separates replay into two decisions: learning value assessment and policy compatibility.
- Demonstrates superior performance over naive replay and competitive results against complex replay methods.
- Achieves efficiency gains in scenarios dominated by environment interaction costs.
Read more
Headroom-Drift Replay: A Primitive for Principled Replay Control in GRPO
Summary
This paper addresses the challenges of reinforcement learning (RL) in post-training for reasoning models, particularly in agentic settings where environment interaction incurs significant costs. The authors introduce 'Headroom-Drift Replay,' a novel replay control primitive designed for Group Reinforcement Policy Optimization (GRPO). This method separates the replay process into two distinct decisions: 'Headroom,' which ranks stored groups based on their remaining learning value, and 'Drift,' which assesses the compatibility of these groups with the current policy. By focusing solely on replay control without additional training machinery, the authors isolate the impact of replay on training dynamics. Empirical evaluations demonstrate that Headroom-Drift Replay outperforms naive replay methods and matches or exceeds more complex replay strategies across various benchmarks, including mathematical reasoning, multimodal reasoning, and Agentic Search. In particular, it shows significant efficiency improvements in Agentic Search scenarios, where it reduces wall-clock time while maintaining quality.
Methodology
The methodology involves the development of Headroom-Drift Replay, which utilizes two key components: Headroom for assessing the learning value of stored groups and Drift for evaluating their compatibility with the current policy. This approach allows for the reuse of entire groups of trajectories rather than individual samples, preserving the structure necessary for effective learning in GRPO. The authors conduct extensive empirical evaluations across multiple reasoning benchmarks to validate the effectiveness of their proposed method.
Results
The results indicate that Headroom-Drift Replay consistently outperforms naive replay methods and achieves comparable or superior performance to more complex replay strategies across mathematical reasoning, multimodal reasoning, and Agentic Search benchmarks. In particular, it demonstrates significant improvements in efficiency, especially in Agentic Search, where it reduces the need for costly fresh environment interactions while maintaining high-quality outputs.
Implications
The findings suggest that principled replay control can significantly enhance the efficiency and effectiveness of RL training in reasoning models, particularly in environments where interaction costs are high. This could lead to more scalable and practical applications of RL in various domains, including natural language processing and multimodal reasoning tasks.
Distilling deep optical flow stereo methods to retrieve dense three-dimensional wind fields
Computer Vision
Efficient ML
Time Series
- Introduces deep optical flow techniques to enhance the retrieval of three-dimensional wind fields from geostationary satellite imagery.
- Develops a single-satellite model that reduces computational costs and improves coverage compared to traditional stereo methods.
- Demonstrates improved performance of stereo winds over operational AMVs, particularly in specific water vapor bands.
- Addresses the circular dependency of traditional AMV height estimation by eliminating reliance on numerical weather prediction (NWP) background states.
Read more
Distilling deep optical flow stereo methods to retrieve dense three-dimensional wind fields
Summary
This paper addresses the challenges of retrieving dense three-dimensional wind fields from geostationary atmospheric motion vectors (AMVs), which traditionally rely on window-based cross-correlation methods that can be computationally expensive and yield inaccurate height estimates. The authors propose a novel approach that utilizes deep optical flow techniques for stereo matching, replacing conventional tracking methods to enhance retrieval efficiency and accuracy. By fine-tuning a model that balances self-supervised geometric residual loss with supervised radiosonde reconstruction, the authors develop a single-satellite student model that emulates the performance of a more complex stereo teacher model. This approach eliminates the need for multi-satellite overlap, allowing for global wind generation from full-disk geostationary imagery. The validation of the model against radiosondes, operational AMVs, and other datasets demonstrates that the stereo winds significantly outperform traditional AMVs, particularly in water vapor bands, although some degradation is observed in long-wave infrared bands. Overall, this work presents a significant advancement in atmospheric wind retrieval methodologies, leveraging machine learning to improve data assimilation systems.
Methodology
The authors replace traditional window-based tracking in stereo matching with deep optical flow techniques. They fine-tune a model that combines self-supervised geometric residual loss with supervised reconstruction from radiosonde data. A single-satellite student model is distilled from a stereo teacher model to generate wind fields globally without requiring multi-satellite overlap. The model's performance is validated using various datasets, including radiosondes and operational AMVs.
Results
The results indicate that the stereo winds generated by the proposed method outperform operational AMVs in terms of accuracy, particularly in the water vapor bands (6.2, 6.9, and 7.3 µm). However, some degradation in performance is noted in the long-wave infrared (11.2 µm) band. The validation through triple collocation confirms the effectiveness of the new approach in producing dense wind fields.
Implications
This research has significant implications for meteorological data assimilation systems, as it provides a more accurate and efficient method for retrieving atmospheric wind fields. The ability to generate high-density wind vectors without reliance on NWP can enhance weather forecasting and improve our understanding of atmospheric dynamics.
Pushing the (Decision) Boundaries: Dynamically Calibrating Differentially Private Noise to Explainability in Federated Learning
Federated Learning
Interpretability
- XCal-FL dynamically calibrates DP noise during training to improve explanation fidelity.
- The method utilizes three complementary signals to adjust noise levels effectively.
- Experiments show significant improvements in both predictive performance and explanation fidelity compared to static-noise approaches.
- The study reveals that explanation fidelity exhibits non-linear dynamics with respect to privacy loss, differing from predictive performance.
Read more
Pushing the (Decision) Boundaries: Dynamically Calibrating Differentially Private Noise to Explainability in Federated Learning
Summary
This paper addresses the challenge of maintaining explanation fidelity in Federated Learning (FL) when incorporating Differential Privacy (DP). The authors propose XCal-FL, a novel training algorithm that dynamically calibrates DP noise based on three signals: prediction logit variations, counterfactual margins, and saliency concentration. This approach allows for a closed-loop system that adapts noise levels during training, enhancing both predictive performance and explanation fidelity. The experiments conducted on medical imaging datasets demonstrate that XCal-FL significantly outperforms static-noise FL methods, achieving over 10% improvement in accuracy and up to 5× improvement in explanation fidelity. The findings indicate that explainability is a distinct dimension of the privacy trade-off, necessitating dedicated strategies for training and privacy-budget allocation in sensitive applications.
Methodology
The authors developed XCal-FL, a closed-loop training algorithm for image classification in cross-silo FL. It dynamically adjusts DP noise based on three signals: prediction logit variations, counterfactual margins, and saliency concentration, while ensuring formal DP guarantees through adaptive privacy accounting.
Results
XCal-FL demonstrated over 10% improvement in predictive performance and up to 5× improvement in explanation fidelity compared to static-noise FL methods. It also achieved higher privacy-budget efficiency, translating cumulative privacy loss into greater gains in accuracy and explanation fidelity.
Implications
The findings suggest that explainability should be treated as a separate dimension in the privacy-utility trade-off, particularly in applications requiring trustworthy model explanations, such as healthcare. This has implications for how privacy budgets are allocated in Federated Learning settings.
Out-of-Distribution Generalisation with Sequence Models in Offline Multi-Agent Reinforcement Learning
Reinforcement Learning
- Task diversity is more critical than dataset size for zero-shot generalization in offline MARL.
- The proposed multi-task approach significantly outperforms single-task models and behavior cloning baselines.
- A new multi-task offline MARL evaluation suite is introduced, enhancing the benchmarking of generalization capabilities.
- Model capacity positively influences generalization performance for challenging tasks.
Read more
Out-of-Distribution Generalisation with Sequence Models in Offline Multi-Agent Reinforcement Learning
Summary
This paper addresses the challenge of generalizing to unseen tasks in offline multi-agent reinforcement learning (MARL). The authors conduct a thorough analysis of zero-shot task generalization and investigate the effects of task diversity, dataset size, and model capacity on performance. They extend offline sequence modeling architectures to accommodate multi-task observation and action spaces with variable agent counts. The key finding is that increasing task diversity is more crucial than merely enlarging the dataset size for achieving robust zero-shot transfer. Through extensive experiments across four environments (Connector, RWARE, SMAX, and LBF), the multi-task approach shows a mean performance improvement of 3.2x on held-out test tasks compared to single-task models, outperforming strong behavior cloning baselines. The work emphasizes the importance of diverse training distributions for developing generalizable MARL agents and provides a roadmap for effectively scaling offline MARL.
Methodology
The authors developed a multi-task offline MARL evaluation suite and conducted large-scale experiments using three state-of-the-art offline multi-agent sequence models. They analyzed the relationship between task diversity, dataset size, and model capacity on zero-shot generalization performance, employing a Decentralized Partially Observable Markov Decision Process (Dec-POMDP) framework for their experiments.
Results
The multi-task models achieved an average improvement of 3.2x in performance on held-out test tasks compared to single-task models. The results indicated that increasing the number of tasks in the training data led to significant enhancements in zero-shot transfer capabilities, while simply increasing dataset size did not yield similar benefits.
Implications
The findings suggest that future research and applications in offline MARL should focus on enhancing task diversity within training datasets to improve generalization to unseen tasks. This could have significant implications for deploying MARL agents in real-world scenarios where data collection is constrained.
The Geometry of Ignorance: LLMs Know When to Temper Bayesian Priors
NLP
Large Language Models
Theory
- The 'direction of ignorance' in LLMs encodes the unigram prior and is universally present across various model families.
- The prior loading factor (λ) effectively measures the reliance on the unigram prior, varying with the informativeness of the context.
- The study provides a tempered Bayesian interpretation of predictions, allowing for meaningful comparisons across models.
- Causal interventions on λ can steer predictions toward or away from the unigram prior, indicating its active role in the prediction process.
Read more
The Geometry of Ignorance: LLMs Know When to Temper Bayesian Priors
Summary
This paper investigates how large language models (LLMs) handle uncertainty in predictions by utilizing a geometric framework. The authors identify a specific direction in the unembedding space, termed the 'direction of ignorance,' which encodes the unigram distribution of the training corpus. This direction serves as a Bayesian prior that the model defaults to when context is insufficient. The study examines four model families (Llama, Qwen, Gemma, and Pythia) and demonstrates that the prior loading factor (λ) decreases as context becomes more informative. The authors provide a geometric-probabilistic interpretation of the prediction state, decomposing it into a prior component and a context-driven likelihood. They establish that this direction is causally active and can be manipulated to influence predictions. The findings suggest that larger models exhibit lower reliance on the prior when provided with rich context, and the paper invites further exploration into the implications of these dynamics across different model architectures.
Methodology
The authors employed a forward-pass-free method to identify the direction of ignorance in the unembedding space through linear least-squares fitting. They analyzed the relationship between the prediction state and the unigram prior, decomposing the prediction into components that reflect both prior and context-driven influences.
Results
The study confirmed the existence and uniqueness of the direction of ignorance across all examined model families. It demonstrated that the prior loading factor λ decreases with increasing context informativeness and that manipulating λ can directly influence model predictions. Additionally, the findings revealed a scaling trend where larger models tend to rely less on the prior in high-context situations.
Implications
These insights could enhance the understanding of LLM behavior in uncertain contexts, potentially informing the design of more robust models that can better leverage context. The findings may also contribute to advancements in interpretability and the calibration of predictions in language models.
Kernel Reboot: Breaking the Boundaries of Neural Tangent Kernels for Neural Fields
Computer Vision
Theory
Efficient ML
- Introduces three algorithms to enhance Neural Fields using Neural Tangent Kernels.
- NTK-KIP enables effective inpainting from sparse data by learning a distilled support set.
- MetaQuill allows fast adaptation to new scenes with minimal task-specific weight adjustments.
- MetaQuill-KIP combines the strengths of both NTK-KIP and MetaQuill for superior performance.
Read more
Kernel Reboot: Breaking the Boundaries of Neural Tangent Kernels for Neural Fields
Summary
This paper addresses the challenges of high-quality reconstruction from sparse observations in Neural Fields (NFs), which map continuous coordinates to signals like color or density. Traditional Neural Tangent Kernel (NTK) regression, while providing closed-form fits, is limited by its linear nature and inability to accumulate reusable task priors. The authors propose three innovative algorithms: NTK-KIP, which learns a distilled support set of coordinates for effective inpainting; MetaQuill, which meta-learns a shared initialization for fast adaptation to new scenes; and MetaQuill-KIP, which combines both approaches for enhanced performance. These methods enable high-PSNR reconstructions and semantically plausible inpainting with minimal adaptation, contrasting with diffusion-style baselines that require extensive pretraining and tuning. The findings demonstrate that NTK-driven neural fields can achieve non-linear and meta-learnable characteristics, bridging the gap between analytical kernels and practical few-shot reconstruction.
Methodology
The authors developed three algorithms: NTK-KIP for learning a compact representation from sparse data, MetaQuill for meta-learning a shared initialization for quick adaptation, and MetaQuill-KIP which integrates both methods. These algorithms leverage the properties of Neural Tangent Kernels to facilitate efficient training and inference, particularly in matrix completion tasks.
Results
MetaQuill-KIP achieved high Peak Signal-to-Noise Ratio (PSNR) reconstructions and semantically plausible inpainting under very sparse observations, requiring only lightweight per-instance adaptation. This performance surpasses that of traditional diffusion-style baselines, which typically rely on large pretrained generative models and extensive tuning.
Implications
The proposed methods can significantly enhance the efficiency and quality of tasks involving Neural Fields, such as image reconstruction and 3D scene representation. This work opens avenues for further research in few-shot learning and representation learning using NTKs, potentially impacting various applications in computer vision and graphics.