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
Internal-State Probes Read the Situation, Not the Action: Three Negative Results for Pre-Action Misalignment Monitoring
NLP
Large Language Models
Interpretability
- Internal-state probes do not reliably predict future misaligned actions.
- The study identifies a recurring failure mode where probes read the situation rather than the action.
- Three distinct probing methods were tested across multiple model families, yielding negative results.
- The paper contributes a methodology for evaluating internal-state probes as pre-action monitors.
Read more
Internal-State Probes Read the Situation, Not the Action: Three Negative Results for Pre-Action Misalignment Monitoring
Summary
This paper investigates the effectiveness of internal-state probes in monitoring agentic systems for potential misalignment before harmful actions are taken. The authors explore whether these probes can reliably predict future misaligned behaviors based on internal model states. They conduct tests across three different model families: Qwen2.5-Coder-32B-Instruct, Llama-3.1-8B-Instruct, and Gemma-3-27B-IT, using various probing techniques. The findings reveal that while some probes can describe the current situation or prompt, they fail to robustly predict future actions due to issues of generalization and specificity. The authors highlight that the probes often read the situation rather than the action, leading to a lack of reliable pre-action monitoring. They provide a methodology for converting internal-readout claims into pre-action tests and report negative results, emphasizing that the tested probe families do not yield a robust pre-action signal under the conditions evaluated.
Methodology
The authors employed three probing techniques: a disposition direction based on difference-of-means vectors, a behavioral prefill probe trained on hidden states, and emotion-concept vectors for steering. Each method was evaluated against specific metrics, including threshold crossing, AUC/lift, and rate contrasts under matched-norm steering, across different model families.
Results
The results indicated that while some probes achieved high accuracy in their construction settings, they failed to generalize to pre-action contexts. For instance, the Qwen model showed perfect separation in training but failed to activate in relevant pre-action scenarios. The Llama model's features decoded prompts effectively but did not predict future behaviors robustly. The Gemma emotion projections were semantically meaningful but lacked specificity in predicting actions.
Implications
The findings suggest that current methodologies for monitoring agentic systems may need to be reevaluated, as existing internal-state probes do not provide reliable pre-action signals. This has implications for the development of safer AI systems, highlighting the need for more robust monitoring techniques that can effectively predict misaligned behaviors.
Randomized neural operator for parametric PDEs with fast training and conformal uncertainty quantification
Efficient ML
Optimization
Theory
- PCA–RaNN combines PCA-based dimensionality reduction with randomized neural networks for fast training.
- The method reduces training time significantly while maintaining accuracy compared to traditional neural operators.
- Ensemble averaging is used to provide conformal prediction intervals for uncertainty quantification.
- The framework allows for rapid online adaptation without retraining hidden features.
Read more
Randomized neural operator for parametric PDEs with fast training and conformal uncertainty quantification
Summary
The paper introduces PCA–RaNN, a novel randomized latent neural operator designed for efficiently solving parametric partial differential equations (PDEs). Traditional neural operators often require extensive non-convex training, which can be computationally expensive and sensitive to various factors. PCA–RaNN addresses these challenges by integrating PCA-based dimensionality reduction with fixed random features and a closed-form least-squares readout, transforming the latent operator learning into a linear regression problem. This approach significantly reduces training time by one to three orders of magnitude while maintaining competitive accuracy across various benchmarks, including Burgers, Darcy, Navier–Stokes, and backward heat equations. The authors also propose an energy-matched scaling rule and a lightweight BFGS refinement to optimize feature scales. Ensemble averaging is employed to reduce predictive variance, and the linear readout allows for rapid online adaptation through recursive least squares without the need for retraining hidden features. Overall, PCA–RaNN serves as an efficient and uncertainty-aware surrogate model suitable for many-query scientific workflows.
Methodology
The PCA–RaNN framework utilizes PCA for dimensionality reduction of input and output fields, replacing the nonlinear latent regressor with a fixed random feature map. The output layer is trained using a closed-form least-squares approach, transforming the problem into linear regression. An energy-matched scaling rule is introduced for feature initialization, complemented by a lightweight BFGS refinement for optimization.
Results
PCA–RaNN demonstrated a significant reduction in training time (1-3 orders of magnitude) while achieving competitive accuracy across multiple PDE benchmarks. The method effectively reduced predictive variance through ensemble averaging and provided reliable uncertainty quantification via split-conformal prediction intervals.
Implications
The PCA–RaNN framework offers a promising approach for real-time prediction and extensive parameter exploration in scientific computing, particularly in applications involving uncertainty quantification, design optimization, and inverse problems. Its efficiency and adaptability make it suitable for many-query workflows in various fields such as fluid dynamics and solid mechanics.
Modification-Considering Value Learning for Reward Hacking Mitigation in RL
Reinforcement Learning
- Introduction of Modification-Considering Value Learning (MCVL) to mitigate reward hacking in RL.
- MCVL evaluates incoming transitions by forecasting two training paths and scoring them with a bootstrapped-return estimator.
- Empirical results show MCVL's effectiveness across multiple environments while maintaining performance.
- Formalization of safety and permissiveness guarantees for MCVL's transition filtering mechanism.
Read more
Modification-Considering Value Learning for Reward Hacking Mitigation in RL
Summary
This paper addresses the issue of reward hacking in reinforcement learning (RL), where agents exploit poorly specified reward signals to achieve high returns without fulfilling the intended objectives. The authors introduce Modification-Considering Value Learning (MCVL), a novel approach that operationalizes the concept of current utility optimization for value-based RL. MCVL wraps an off-policy learner and evaluates incoming transitions as potential modifications by forecasting two training paths: one that includes the transition and one that does not. Both paths are scored using a frozen bootstrapped-return estimator derived from a learned reward model and value function. A transition is accepted only if its inclusion does not lower the score. The method is instantiated with Double Deep Q-Network (DDQN) and Twin Delayed Deep Deterministic Policy Gradient (TD3) algorithms. The authors empirically demonstrate that MCVL effectively mitigates reward hacking across various safety-relevant gridworlds and modified MuJoCo tasks while maintaining performance comparable to an Oracle trained on the true reward. Additionally, the paper formalizes safety and permissiveness guarantees for MCVL's gating rule, emphasizing the importance of an accurate return estimator. The findings suggest that MCVL can be a practical safeguard for RL systems, particularly in safety-critical applications.
Methodology
The methodology involves wrapping an off-policy value-based RL learner with MCVL, which treats each incoming transition as a candidate modification. For each transition, MCVL forecasts two training scenarios (with and without the transition) and scores both using a frozen bootstrapped-return estimator derived from a learned reward model and value function. The transition is accepted only if it does not decrease the score, thus mitigating potential reward hacking.
Results
MCVL was empirically tested across four safety-relevant gridworlds and three modified MuJoCo continuous-control tasks. The results indicate that MCVL effectively mitigates reward hacking while achieving performance levels comparable to an Oracle trained on the true reward. The method demonstrated robustness across diverse hacking mechanisms, confirming its applicability in various environments.
Implications
The findings suggest that MCVL can serve as a practical safeguard for reinforcement learning systems, particularly in safety-critical applications such as autonomous driving and medical diagnostics. By effectively addressing reward hacking, MCVL enhances the reliability and safety of RL agents in real-world scenarios.
Singular Learning and Occam's Razor in Deep Monomial Networks
Theory
- Critical points in deep monomial networks correspond to subnetworks with inactive or redundant neurons.
- The study provides a mathematical perspective on the implicit bias towards simplicity in deep learning models.
- Mason's Theorem is utilized to analyze the rank-deficiency of the Jacobian in the context of neural networks.
- The findings align with the principle of Occam's Razor, suggesting that neural networks tend to converge to simpler functions.
Read more
Singular Learning and Occam's Razor in Deep Monomial Networks
Summary
This paper investigates the critical points of deep fully-connected networks with monomial activation functions through the lens of Singular Learning Theory (SLT). The authors demonstrate that these critical points, where the Jacobian of the model's parametrization is rank-deficient, correspond to subnetworks where certain neurons are inactive or redundant. This finding provides a mathematical foundation for understanding the implicit bias in deep learning models, particularly their tendency to converge towards simpler functions, akin to the principle of Occam's Razor. The study employs polynomial algebra tools, specifically Mason's Theorem, to analyze the critical points and their implications for the learning dynamics of these networks. The results suggest that the architecture of deep monomial networks inherently leads to a preference for simpler configurations during training, reinforcing the notion of implicit biases in neural network optimization.
Methodology
The authors utilize tools from polynomial algebra, particularly Mason's Theorem, to analyze the critical points of deep fully-connected networks with monomial activations. They focus on the rank-deficiency of the Jacobian of the model's parametrization and employ a divisibility analysis of the polynomials defined by the network.
Results
The main result indicates that for sufficiently large activation degrees, the critical points of the parametrization are precisely the subnetworks where some neurons can be removed without altering the function realized by the network. This finding highlights the relationship between critical points and the implicit biases in learning, suggesting that deep monomial networks inherently favor simpler configurations.
Implications
The results have implications for understanding the optimization dynamics of deep learning models, particularly in how they converge to simpler solutions. This insight can inform the design of neural network architectures and training algorithms that leverage the implicit biases towards sparsity and simplicity.
COOPA: A Modular LLM Agent Architecture for Operations Research Problems
Large Language Models
Optimization
Interpretability
- COOPA improves formulation quality and accuracy in OR decision-making.
- The architecture provides traceability and confidence explanations for model elements.
- It supports multiple solver backends, enhancing adaptability for different OR problems.
- Empirical results show COOPA outperforms existing LLM-based OR systems.
Read more
COOPA: A Modular LLM Agent Architecture for Operations Research Problems
Summary
The paper introduces COOPA (COoperative OPerations Agent), a modular architecture designed to enhance decision support in Operations Research (OR) through the use of large language models (LLMs). Traditional OR modeling is complex and requires significant expertise, often leading to underutilization in organizations. COOPA addresses three main limitations of existing LLM-based systems: low accuracy on complex problems, opaque outputs, and limited solver adaptability. The architecture consists of three key components: iterative confidence-based modeling, which generates and evaluates multiple candidate formulations to select the best one; element-level provenance and confidence explanations, which provide traceability from problem descriptions to model formulations; and multi-solver routing that directs problems to specialized optimizer agents. The authors demonstrate that COOPA outperforms existing methods across various benchmarks and LLM backbones, achieving improved accuracy and providing interpretable outputs that facilitate human verification.
Methodology
COOPA employs a modular architecture that integrates iterative confidence-based modeling to generate and evaluate multiple formulations, provides element-level provenance for transparency, and implements a multi-solver routing system to direct problems to appropriate optimizer agents based on problem type.
Results
COOPA achieved the best macro-average accuracy on six out of eight LLM backbones tested and improved over the strongest baseline by up to 6.7 percentage points. The ablation studies highlighted the significant contribution of the iterative confidence-based modeling component.
Implications
The COOPA architecture has the potential to democratize access to OR modeling by reducing the expertise barrier, making it easier for organizations to implement effective decision-making processes. Its modular design allows for future enhancements and adaptations to various OR problems and solver technologies.
Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders
Multimodal
- Introduces the concept of cross-modal feature heterogeneity, highlighting discrepancies in feature directions for the same concept across modalities.
- Demonstrates that existing alignment approaches may degrade feature recovery by collapsing distinct feature directions into a single coordinate.
- Proposes a method using modality-specific sparse autoencoders to preserve feature geometry and align features post hoc.
- Shows improvements in reconstruction fidelity and cross-modal retrieval performance with the proposed method.
Read more
Same Concept, Different Directions: Cross-Modal Feature Heterogeneity in Sparse Autoencoders
Summary
This paper addresses the limitations of vision-language models (VLMs) that map images and text into a joint embedding space, where embeddings often entangle multiple semantic features, complicating their interpretability. The authors challenge the assumption that semantically corresponding features share the same directions across modalities, introducing the concept of cross-modal feature heterogeneity. They demonstrate that this heterogeneity leads to a modality split, where the same concept activates different latent representations depending on the modality. The paper proposes a novel approach that employs modality-specific sparse autoencoders (SAEs) to maintain the unique feature geometry of each modality, followed by a post hoc alignment of corresponding features. This method enhances reconstruction fidelity and improves performance in cross-modal retrieval and concept steering, providing a more effective framework for interpreting multimodal embeddings.
Methodology
The authors utilize sparse autoencoders to analyze joint embeddings from vision-language models, focusing on the feature directions of corresponding concepts across image and text modalities. They propose a new training approach that maintains modality-specific feature geometries and aligns features after training, rather than during the training process.
Results
The proposed approach significantly improves reconstruction fidelity and enhances performance in cross-modal retrieval tasks. It also facilitates more effective concept steering, demonstrating that preserving the unique feature geometry of each modality leads to better interpretability and usability of the embeddings.
Implications
The findings suggest that addressing cross-modal feature heterogeneity is crucial for improving the interpretability and effectiveness of multimodal models. This work has potential applications in enhancing the performance of vision-language models in various tasks, including image captioning, visual question answering, and cross-modal retrieval.
Adaptive Financial Transformer with Regime-Gated Attention for Stock Return Prediction
Time Series
- Introduction of the Adaptive Financial Transformer (AFT) for stock return prediction.
- Dynamic biasing of self-attention weights based on market regimes and feature similarities.
- Correction of backtesting leakage issues in prior literature.
- Development of a Financially-Aware Composite Objective to enhance prediction accuracy.
Read more
Adaptive Financial Transformer with Regime-Gated Attention for Stock Return Prediction
Summary
This paper presents the Adaptive Financial Transformer (AFT), a novel deep learning architecture designed for stock return prediction, addressing the challenges posed by low signal-to-noise ratios and non-stationary market regimes. Traditional sequence models often struggle with overfitting and noise propagation, particularly in financial contexts where the relationships between indicators are complex. The AFT incorporates a Market Regime Encoder, an Adaptive Gate Network, and an Adaptive Financial Context module to enhance self-attention mechanisms by dynamically adjusting attention weights based on the semantic relationships of financial indicators. The model segments 95 technical features into 11 categories, utilizing an unsupervised approach to project market regimes and scale similarity matrices. Additionally, the paper identifies and corrects a significant backtesting leakage found in previous literature, which inflated performance metrics. To improve prediction accuracy, a Financially-Aware Composite Objective is introduced, combining Mean Absolute Error, directional sign accuracy, and non-overlapping Sharpe ratios. The AFT is rigorously evaluated against standard machine learning and recurrent models, demonstrating competitive performance while reducing model complexity by 15.2% and utilizing a more efficient feature subset.
Methodology
The AFT architecture employs a Market Regime Encoder to categorize financial indicators, an Adaptive Gate Network to dynamically adjust attention weights, and an Adaptive Financial Context module to enhance the model's predictive capabilities. The model is evaluated using paired t-tests and Cohen’s d effect sizes across multiple random seeds.
Results
The AFT model demonstrated comparable predictive performance to standard baselines while reducing the parameter footprint by 15.2%, achieving a more efficient representation with a Top-40 feature subset. The introduction of a Financially-Aware Composite Objective improved the model's ability to avoid regression-to-the-mean issues.
Implications
The findings suggest that incorporating regime-based attention mechanisms can significantly enhance stock return prediction models, potentially leading to better investment strategies and risk management in quantitative finance.
OverFlowLight: Real-Time Gridlock Prevention and Traffic Signal Optimization for Urban Intersections
Reinforcement Learning
Optimization
Computer Vision
- OverFlowLight effectively detects and mitigates traffic overflow in real-time.
- The framework integrates multi-modal sensing and reinforcement learning for enhanced TSC performance.
- Real-world implementations show a 60.4% reduction in overflow incidents and an 18.2% increase in throughput.
- The system minimizes the reliance on manual traffic control interventions.
Read more
OverFlowLight: Real-Time Gridlock Prevention and Traffic Signal Optimization for Urban Intersections
Summary
The paper presents OverFlowLight, a novel framework aimed at preventing traffic gridlock and optimizing traffic signal control (TSC) at urban intersections. Traditional TSC algorithms often prioritize throughput but fail to address queue overflow, especially during peak traffic hours, leading to severe congestion and safety hazards. OverFlowLight introduces a real-time overflow detection mechanism utilizing multi-modal sensing from cameras and radars to identify overflow conditions. Upon detection, it dynamically generates dedicated overflow phases in the signal cycle to alleviate blocking queues. The framework employs a hybrid control design that combines rapid rule-based interventions with reinforcement learning (RL) for long-term efficiency. Extensive real-world deployments across 43 intersections in three major cities demonstrate OverFlowLight's effectiveness, reducing overflow incidents by 60.4% and increasing network throughput by 18.2% compared to existing baselines. The system's modularity allows seamless integration with existing RL-based TSC agents, significantly reducing the need for manual interventions typically required in expert-tuned signal plans. This work represents a significant advancement in developing scalable, data-driven solutions for urban traffic management, contributing to more resilient and efficient transportation systems.
Methodology
The methodology involves a three-stage pipeline: (1) real-time overflow detection using radar and camera data, (2) construction of overflow-clearing phases based on detected overflow directions, and (3) application of these phases through either traditional or RL-based controllers for immediate intervention.
Results
The empirical results indicate that OverFlowLight reduces overflow incidents by 60.4% and increases network throughput by 18.2% compared to deployed baseline systems, showcasing its effectiveness in real-world traffic scenarios.
Implications
The implications of this research extend to urban traffic management, where OverFlowLight can significantly improve traffic flow, reduce congestion-related accidents, and enhance the overall efficiency of transportation systems in urban environments.
Deployment-Side Adaptiveness in Multi-Horizon Volatility Forecasting
Time Series
- A trained MIMO forecaster can induce a family of deployable predictors through different inference-time rollout rules.
- Non-default rollout rules often outperform standard MIMO deployment, but the best rule varies by architecture and horizon.
- Validation-based deployment policies can provide significant improvements in predictive performance at lower costs.
- Deployment choices are sensitive to the evaluation metric used, affecting the transferability of policies across different loss functions.
Read more
Deployment-Side Adaptiveness in Multi-Horizon Volatility Forecasting
Summary
This paper investigates the impact of deployment strategies on the performance of multi-horizon volatility forecasting models. The authors argue that a trained multi-output (MIMO) forecaster can generate a family of forecasts through different inference-time rollout rules, rather than being limited to a single default deployment. By analyzing 20 stock-volatility series across various architectures and forecast horizons, the study finds that non-default rollout rules often yield better accuracy and cost profiles compared to standard MIMO deployment. However, the optimal rollout rule is architecture- and horizon-dependent, suggesting that a static deployment strategy may not be effective. The authors propose validation-based deployment policies that leverage this variability, demonstrating that validation-selected singletons can enhance performance at lower costs, while small subsets of rules can approximate the benefits of larger ensembles. Furthermore, the study highlights the metric sensitivity of deployment choices, noting that policies optimized for mean squared error (MSE) do not necessarily perform well under alternative metrics like QLIKE. The findings underscore the importance of considering deployment strategies in financial forecasting, advocating for a more nuanced evaluation of trained models based on their operational deployment policies.
Methodology
The authors conducted experiments on 20 stock-volatility series using various forecasting architectures, including linear models and PatchTST. They evaluated the performance of different inference-time rollout rules and employed validation-based deployment policies to optimize forecast accuracy and cost. The study compared the effectiveness of these policies under different metrics, particularly MSE and QLIKE.
Results
The study found that non-default rollout rules generally improved forecasting accuracy compared to the default MIMO approach. Validation-selected singletons provided low-cost enhancements, while small subsets of rules could recover much of the performance benefits of larger ensembles. The results also indicated that the effectiveness of deployment policies is metric-sensitive, with MSE-optimized policies not transferring uniformly to QLIKE.
Implications
The findings suggest that financial forecasting models should be evaluated not only based on their architecture but also on their deployment strategies. This approach could lead to more efficient forecasting systems that better meet operational constraints in real-world applications.
PairSAE: Mechanistic Interpretability from Pair Representations in Protein Co-Folding
Interpretability
- Introduction of PairSAE to improve interpretability in protein co-folding models.
- Utilization of N-mode SVD to summarize pairwise representations effectively.
- Demonstrated alignment of extracted features with biological annotations.
- Enhanced prediction of Boltz-2 affinity values using the proposed method.
Read more
PairSAE: Mechanistic Interpretability from Pair Representations in Protein Co-Folding
Summary
The paper introduces PairSAE, a novel approach to enhance mechanistic interpretability in protein co-folding by leveraging pairwise representations. Traditional sparse autoencoders (SAEs) struggle with pairformer architectures due to the quadratic increase in feature dimensions when processing pairwise data. PairSAE addresses this by employing an N-mode singular value decomposition (SVD) to summarize pairwise tensors into token-wise interaction roles, allowing for a shared set of token-level features that can decode both sequence and pair representations. The method is evaluated using Boltz-2 activations for protein-ligand complexes, demonstrating that PairSAE can extract interpretable features that align with UniProt annotations and effectively predict Boltz-2 affinity values. This approach not only clarifies the latent space of foundation models in structural biology but also mitigates the pitfalls associated with conventional SAEs in the context of pairformer architectures.
Methodology
PairSAE employs N-mode singular value decomposition (SVD) to compress pairwise tensor representations into a more manageable form, allowing for the extraction of token-wise interaction roles. It then utilizes a sparse autoencoder to learn a shared set of features that can reconstruct both sequence-level and pairwise embeddings, facilitating interpretability.
Results
The application of PairSAE to Boltz-2 activations resulted in interpretable features that closely matched UniProt annotations. The model also achieved improved predictions of Boltz-2 affinity values, indicating its effectiveness in linking latent representations to meaningful structural concepts.
Implications
PairSAE has the potential to advance the field of structural biology by providing clearer insights into the mechanisms of protein folding and interactions. This could enhance the reliability of predictions made by foundation models and facilitate the design of new proteins and small molecules.
Adaptive Block Diffusion: Resolving Training-Inference Mismatch in Diffusion Language Models
NLP
Large Language Models
Generative Models
- Introduction of Adaptive Block Diffusion (ABD) to resolve training-inference mismatch in DLMs.
- ABD optimizes denoising risk over a distribution of prefix-window configurations, enhancing generalization.
- The framework guarantees denoising optimality for any inference policy supported during training.
- Empirical results show ABD's structural invariance and superior performance compared to fixed-block models.
Read more
Adaptive Block Diffusion: Resolving Training-Inference Mismatch in Diffusion Language Models
Summary
This paper introduces Adaptive Block Diffusion (ABD), a novel framework designed to address the training-inference mismatch prevalent in Diffusion Language Models (DLMs). Traditional DLMs are constrained by fixed context structures during training, which limits their ability to generalize to arbitrary configurations during inference. This mismatch often results in significant performance degradation when models are evaluated outside their training support. ABD resolves this issue by treating denoising configurations as stochastic variables, allowing the model to optimize denoising risk across a diverse range of prefix-window configurations. The authors demonstrate that ABD guarantees denoising optimality for any inference policy supported during training, thus ensuring structural invariance across different decoding scales. Empirical results show that ABD not only avoids off-grid degradation but also maintains a consistent relationship between block size and perplexity, outperforming fixed-block models at their target configurations while exhibiting strong performance across all scales.
Methodology
The methodology involves training a single model over the full configuration space of prefix-window pairs, treating configurations as stochastic variables. This approach allows for optimization of denoising risk across various configurations, ensuring that the model learns to generalize effectively to unseen configurations during inference.
Results
The results indicate that ABD eliminates off-grid degradation and recovers a monotonic relationship between block size and perplexity. The model matches or outperforms fixed-block specialists at their target configurations while maintaining robust performance across all decoding scales.
Implications
The implications of this work suggest that ABD could significantly enhance the performance of DLMs in practical applications, particularly in scenarios requiring flexible and efficient text generation across varying contexts. This could lead to advancements in natural language processing tasks where inference configurations are not predetermined.
PEBS: Per-rater Empirical-Bayes Shrinkage for RLHF Reward-Model Calibration
Reinforcement Learning
- PEBS offers a solution to the limitations of traditional global reward models in RLHF by focusing on individual annotator calibration.
- The method utilizes empirical-Bayes shrinkage to adjust per-rater calibrators based on population statistics.
- PEBS demonstrated an 8.58% RMSE reduction on the PRISM dataset and a 9.66% RMSE reduction on the PluriHarms dataset.
- The approach maintains the integrity of the base reward model while enhancing the accuracy of individual annotator predictions.
Read more
PEBS: Per-rater Empirical-Bayes Shrinkage for RLHF Reward-Model Calibration
Summary
The paper introduces PEBS, a novel per-rater empirical-Bayes shrinkage estimator designed to improve the calibration of reward models in Reinforcement Learning from Human Feedback (RLHF). Traditional methods aggregate preferences from multiple annotators into a single global model, which can obscure individual differences in rating scales and lead to inaccuracies. PEBS addresses this by fitting individual affine calibrators for each annotator based on a subset of their ratings and applying Morris–James–Stein shrinkage to align these calibrators with the population mean. This approach allows for more accurate representation of individual annotator preferences without retraining the underlying reward model. The effectiveness of PEBS is demonstrated through experiments on two datasets: PRISM and PluriHarms, where it achieved significant reductions in root mean square error (RMSE) compared to the pooled population-slope baseline, indicating improved per-user accuracy in reward modeling.
Methodology
PEBS employs a per-rater empirical-Bayes shrinkage technique, fitting individual affine calibrators for each annotator using a held-out portion of their ratings. It then applies Morris–James–Stein shrinkage to these calibrators to align them with the population mean, allowing for a more accurate representation of individual preferences without the need for retraining the reward model.
Results
The application of PEBS resulted in an 8.58% reduction in RMSE on the PRISM dataset and a 9.66% reduction on the PluriHarms dataset when compared to a pooled population-slope baseline, indicating significant improvements in the accuracy of reward modeling for individual annotators.
Implications
The findings suggest that PEBS can enhance the performance of RLHF systems by providing more accurate reward models that reflect individual annotator preferences. This could lead to better alignment of AI systems with human values and preferences, improving the overall effectiveness of RLHF applications.
Priced Motion Through Optimal Faces: A Normal-Fan Geometry for Non-Stationary Adversarial MDPs
Reinforcement Learning
Theory
Optimization
- Introduces the concept of face-crossing price to quantify regret in non-stationary adversarial MDPs.
- Dynamic regret can be decomposed into face-crossing price and within-face selection error.
- Demonstrates that large loss variations can incur zero regret under certain conditions.
- Establishes that early crossings of decision boundaries are more costly than late ones.
Read more
Priced Motion Through Optimal Faces: A Normal-Fan Geometry for Non-Stationary Adversarial MDPs
Summary
This paper addresses the challenges of non-stationarity in adversarial Markov Decision Processes (MDPs) by introducing a normal-fan geometry framework. Traditional analyses equate the cost of non-stationarity with the magnitude of loss movement, which can be misleading. The author proposes that the occupancy measures of policies form a polytope, with each loss vector revealing an optimal face of this polytope. The concept of 'face-crossing price' is introduced, representing the minimum regret incurred when a learner remains on the previous optimal face despite changes in loss. The paper demonstrates that dynamic regret can be decomposed into this face-crossing price and a within-face selection error. The findings indicate that significant loss variation can occur without incurring regret, and crossing decision boundaries can lead to substantial costs. The methodology includes a geometric interpretation of decision regions and a combinatorial approach to pricing loss motion, which is simplified through a single value backup that computes the expected optimal Bellman advantage. The paper concludes with empirical tests on three benchmarks to validate the theoretical framework.
Methodology
The paper employs a geometric approach to model the occupancy measures of policies as a polytope, utilizing normal-fan geometry to analyze the effects of non-stationarity in adversarial MDPs. The analysis includes defining the face-crossing price and deriving the relationship between dynamic regret and this price through theoretical proofs.
Results
The author proves that for learners tracking the previous optimal face, dynamic regret can be exactly decomposed into the face-crossing price and selection error. The findings also show that loss variation can be large without incurring regret, and that the timing of decision boundary crossings significantly affects regret costs.
Implications
This work has implications for designing more robust reinforcement learning algorithms that can adapt to non-stationary environments, particularly in applications like recommender systems where user preferences change over time. The geometric insights could also inform the development of new algorithms that efficiently manage regret in dynamic settings.
Beyond IID: How General Are Tabular Foundation Models, Really?
Theory
Optimization
Efficient ML
- Introduction of BeyondArena, a unified benchmark for evaluating tabular foundation models across diverse tasks and datasets.
- Development of DataFoundry, a framework for curating high-quality tabular datasets for reproducible research.
- Demonstration that existing TFMs excel on IID tasks but underperform on non-IID and complex datasets compared to traditional models.
- Emphasis on the importance of appropriate data splits and preprocessing for accurate model evaluation.
Read more
Beyond IID: How General Are Tabular Foundation Models, Really?
Summary
This paper addresses the limitations of current evaluations of tabular foundation models (TFMs) in predictive machine learning, which primarily focus on IID (independent and identically distributed) tasks. The authors introduce BeyondArena, a comprehensive benchmark designed to evaluate TFMs across a variety of task types, including IID, temporal, and grouped scenarios, while also considering diverse sample sizes and feature types. They curate 142 datasets from various disciplines and provide a Python framework, DataFoundry, for reproducible dataset curation. The study reveals that while TFMs perform well on small to medium-sized IID datasets, they struggle with non-IID, large-scale, and high-dimensional datasets compared to traditional models. The authors aim to bridge the gap between academic evaluations and real-world applications, emphasizing the need for standardized benchmarking to foster meaningful progress in the field.
Methodology
The authors manually curated 142 datasets following rigorous protocols and developed the BeyondArena benchmark to evaluate TFMs against traditional models across various task types. They utilized a systematic approach to assess model performance, focusing on both IID and non-IID scenarios, and provided a metadata schema for dataset curation through DataFoundry.
Results
The results indicated that state-of-the-art TFMs performed well on tiny to medium-sized IID datasets but failed to compete with traditional tree-based and deep learning models on non-IID, large-scale, and high-dimensional datasets. The study highlighted the necessity of appropriate data handling techniques to ensure valid evaluations.
Implications
The findings suggest that while TFMs have potential, their current limitations in non-IID contexts need to be addressed for practical applications. The introduction of a unified benchmarking framework could accelerate research and development in tabular predictive modeling, ultimately leading to more robust and generalizable models in real-world scenarios.
HybridCodec: Modeling Discrete and Continuous Representations for Efficient Speech Language Models
Audio & Speech
Multimodal
Large Language Models
- HybridCodec integrates discrete tokens with continuous residuals to mitigate information loss in speech models.
- The framework employs a hybrid Transformer architecture for efficient autoregressive and non-autoregressive processing.
- Experimental results show significant improvements in speaker characteristic retention compared to discrete-only methods.
- The approach reduces the number of autoregressive steps needed, enhancing inference efficiency.
Read more
HybridCodec: Modeling Discrete and Continuous Representations for Efficient Speech Language Models
Summary
The paper introduces HybridCodec, a novel framework that combines discrete audio representations with continuous residuals to enhance the performance of speech language models. Traditional discrete audio models often suffer from information loss during discretization, leading to degraded performance in downstream tasks. HybridCodec addresses this issue by employing a hybridized discrete-continuous focal modulation codec and a Transformer architecture, enabling autoregressive inference in the discrete domain while allowing for non-autoregressive prediction and continuous residual upsampling. This approach not only improves the retention of speaker characteristics but also reduces the number of autoregressive steps required during inference. The authors demonstrate that their method significantly outperforms existing discrete-only models, particularly at low frame rates, thus providing a unified framework for various speech tasks such as automatic speech recognition (ASR) and text-to-speech (TTS).
Methodology
The authors developed HybridCodec, which extends the FocalCodec architecture by jointly extracting temporally compressed discrete tokens and modeling remaining information as dimensionality-reduced continuous residuals. The HybridLM Transformer processes these hybrid representations, combining autoregressive prediction for discrete tokens with non-autoregressive prediction and continuous residual upsampling.
Results
Experimental evaluations on the LibriTTS dataset indicate that HybridCodec significantly outperforms discrete-only baselines, especially at low frame rates (e.g., 6.25 Hz), while also reducing the number of autoregressive steps required for inference.
Implications
The proposed HybridCodec framework has the potential to enhance various speech applications, including ASR and TTS, by effectively integrating discrete and continuous representations. This could lead to more efficient and high-fidelity speech synthesis and recognition systems, making it applicable in real-world scenarios such as virtual assistants, automated transcription services, and multimodal dialogue systems.
GNBAN: Graph Neural Basis Attention Networks for Long-Horizon Forecasting over Large Entity Sets
Graph Learning
Time Series
Interpretability
- GNBAN combines graph-based representation learning with interpretable basis-style forecast decomposition.
- The model decomposes forecasts into trend, seasonal, and residual components, enhancing interpretability.
- GNBAN improves forecasting accuracy by 4-5% over baseline models on large-scale retail datasets.
- The architecture supports scalable forecasting across extensive product-store catalogs.
Read more
GNBAN: Graph Neural Basis Attention Networks for Long-Horizon Forecasting over Large Entity Sets
Summary
The paper introduces GNBAN (Graph Neural Basis Attention Network), a novel architecture designed for long-horizon demand forecasting in large-scale retail environments. Traditional forecasting methods struggle to manage the complexity of predicting numerous correlated time series across products and stores. GNBAN addresses this by leveraging a heterogeneous graph representation of retail data, allowing the model to learn shared demand dynamics across a vast catalog. The architecture features a basis-attention forecasting head that decomposes forecasts into three interpretable components: trend, seasonal, and generic. This decomposition enhances interpretability and allows the model to specialize in capturing distinct temporal patterns. The authors evaluate GNBAN on two significant retail forecasting benchmarks, M5 Walmart and Favorita Grocery Sales, demonstrating that it outperforms baseline models by improving volume-weighted WRMSSE by 4-5%. The results indicate that GNBAN effectively combines scalability, predictive accuracy, and interpretability in a unified framework, making it a promising solution for complex retail forecasting challenges.
Methodology
GNBAN employs a heterogeneous graph neural network to represent retail data, capturing relationships between products, stores, and categories. The forecasting head utilizes a basis-attention mechanism, allowing each basis function to learn independently from the historical data of the forecast entity, resulting in a decomposition of forecasts into trend, seasonal, and residual components.
Results
GNBAN demonstrated a 4-5% improvement in volume-weighted WRMSSE compared to baseline graph forecasting models on the M5 Walmart and Favorita Grocery Sales datasets. The qualitative analysis revealed that the model's decomposition provides clear insights into demand drivers, enhancing interpretability without the need for post-hoc explanation methods.
Implications
The GNBAN architecture has significant implications for retail demand forecasting, enabling practitioners to manage large-scale forecasting tasks more effectively while gaining insights into demand patterns. Its ability to provide interpretable forecasts can enhance decision-making processes in retail operations.
Regularized Reward-Punishment Reinforcement Learning
Reinforcement Learning
Robotics
- Introduction of KL-Coupled Policy Regularization (KCPR) for enhanced policy interaction.
- Development of KL-Coupled Soft Optimality (KCSO) and its deep realization, klDMP.
- Implementation of stabilization mechanisms for improved learning dynamics.
- Demonstration of superior performance in safety and stability in robotic navigation tasks.
Read more
Regularized Reward-Punishment Reinforcement Learning
Summary
This paper introduces KL-Coupled Policy Regularization (KCPR), a novel framework for Reward-Punishment Reinforcement Learning (RPRL) that enhances the interaction between reward-seeking and punishment-related policies. Unlike traditional RPRL methods that optimize these policies independently, KCPR allows them to influence each other dynamically, treating each policy as a learned prior for the other. The authors derive KL-Coupled Soft Optimality (KCSO) from KCPR, which leads to coupled soft-optimal policies and KL-regularized Bellman operators. This framework facilitates joint value propagation influenced by both reward and punishment signals. To enhance learning stability, the authors introduce a companion-prior softening mechanism and evaluate various replay-buffer designs to balance experiences related to rewards and punishments. Experimental results in grid-world and robotic navigation tasks demonstrate that the proposed klDMP (a deep realization of KCSO) outperforms traditional methods like DQN, SQL, and soft-DMP in terms of safety, learning stability, and competitive task performance. The findings suggest that policy-level coordination is crucial for integrating multiple behavioral objectives in reinforcement learning systems, particularly in environments where both rewards and punishments are significant.
Methodology
The authors propose KCPR as a regularization principle that allows reward and punishment policies to mutually influence each other through KL divergence. They derive KCSO as a soft-optimality framework and implement klDMP as a practical application of KCSO. The methodology includes a companion-prior softening mechanism and a tailored replay-buffer design to optimize learning from both reward and punishment experiences.
Results
Experiments show that klDMP significantly improves safety and learning stability while maintaining competitive performance in grid-world and robotic navigation tasks compared to DQN, SQL, and soft-DMP. The results indicate that the proposed framework effectively balances efficiency and safety in reinforcement learning.
Implications
The findings suggest that policy-level coordination can enhance the integration of multiple motivational processes in reinforcement learning systems, which may lead to more robust and adaptive AI agents. This approach could be applied to various domains, including robotics, autonomous systems, and complex decision-making scenarios.
Blackknife: Hard-Label Query-Limited Black-Box Attacks on Heterogeneous Graph Neural Networks
Graph Learning
- Blackknife operates under strict black-box conditions with limited access to model information.
- The framework uses local structural information to construct a surrogate model for effective perturbation generation.
- Blackknife demonstrates high attack success rates on benchmark datasets against HGNNs.
- The method remains effective against topology-based defense strategies.
Read more
Blackknife: Hard-Label Query-Limited Black-Box Attacks on Heterogeneous Graph Neural Networks
Summary
The paper introduces Blackknife, a novel framework for conducting hard-label, query-limited, and structure-limited black-box attacks on heterogeneous graph neural networks (HGNNs). Existing adversarial attack methods often rely on access to model gradients, soft prediction scores, or complete graph structures, which are typically unavailable in real-world applications. Blackknife addresses this gap by assuming no access to the victim model's architecture or parameters and only utilizing locally observable one-hop heterogeneous structures along with a limited number of hard-label queries. The framework constructs a local relation-aware surrogate model from observable neighborhoods and optimizes perturbations through projected gradient descent. The optimized perturbations are then discretized into relation-preserving structural rewiring operations, verified using limited hard-label feedback from the victim model. Extensive experiments on benchmark datasets (ACM, DBLP, IMDB) demonstrate that Blackknife achieves high attack success rates against various HGNN models, even under topology-based defenses, highlighting the vulnerability of HGNNs to local structure-limited black-box attacks.
Methodology
Blackknife constructs a local relation-aware surrogate model based on observable one-hop heterogeneous structures. It optimizes perturbations by relaxing discrete edge modifications into continuous soft weights, which are then optimized using projected gradient descent. The final perturbations are discretized into structural rewiring operations verified through limited hard-label queries to the victim model.
Results
The experiments show that Blackknife consistently achieves strong attack success rates across three benchmark datasets (ACM, DBLP, IMDB), indicating its effectiveness even when the victim model is defended by topology-based strategies.
Implications
The study highlights the need for improved defenses against adversarial attacks in HGNNs, particularly in high-stakes applications such as financial analysis and biological data processing. It also suggests that current models may require reevaluation of their robustness in real-world deployment scenarios.
Applicability of memorization indicators for early spotting of overfitting while recalibrating sEMG-decoders on low sample sizes
Theory
Robotics
Efficient ML
- Memorization indicators based on ReLU activation statistics can detect overfitting in low-sample sEMG calibration.
- Traditional overfitting metrics are impractical in scenarios with limited data, necessitating alternative approaches.
- The study demonstrates a correlation between test accuracy drops and changes in activation rates during fine-tuning.
- Transfer learning is utilized to improve calibration speed and accuracy for individual users in sEMG applications.
Read more
Applicability of memorization indicators for early spotting of overfitting while recalibrating sEMG-decoders on low sample sizes
Summary
This paper addresses the challenge of overfitting in deep learning models for surface electromyography (sEMG) during user-specific calibration, particularly when sample sizes are limited. The authors explore the use of memorization indicators based on the activation statistics of rectified linear units (ReLU) in deep neural networks (DNNs) as a means to detect overfitting early in the calibration process. Traditional overfitting indicators, such as validation performance metrics, are often impractical in low-sample scenarios due to the lack of additional held-out data. The study conducts a transfer-learning experiment on a benchmark sEMG dataset, where a convolutional neural network is pre-trained on multiple subjects and fine-tuned on individual users with minimal repetitions. The authors monitor both decoding performance and the activation behavior of the last hidden layer during calibration. The findings suggest that decreases in test accuracy during fine-tuning correlate with specific changes in activation rates, indicating that activation-based memorization indicators can effectively signal unsuccessful learning in low-sample sEMG calibration settings. This work highlights the potential of these indicators to enhance the calibration process of sEMG decoders, making them more user-friendly and effective in practical applications.
Methodology
The authors implemented a transfer-learning experiment using a convolutional neural network. The model was pre-trained on a diverse set of subjects and then fine-tuned on individual users with a limited number of repetitions. During the calibration phase, they monitored decoding performance and the activation behavior of the last hidden layer, specifically focusing on the activation rates of ReLU units to identify signs of overfitting.
Results
The results indicate that a decline in test accuracy during the fine-tuning process is accompanied by a significant reduction in activation rates. This correlation suggests that memorization indicators based on activation statistics can serve as effective tools for early detection of overfitting in low-sample sEMG calibration scenarios. The findings were consistent across experiments involving data from 10 subjects.
Implications
The study's findings have important implications for the development of user-friendly sEMG decoders, particularly in medical and assistive technology applications. By utilizing memorization indicators, developers can enhance the calibration process, ensuring that models remain effective even with limited training data. This could lead to broader adoption of sEMG-based systems in real-world settings, such as rehabilitation and human-machine interaction.
CAREBench: A Child-Safety Risk Benchmark for Language Models
NLP
Large Language Models
- CAREBench evaluates upstream child-safety risks in language models, beyond just explicit abuse content.
- The benchmark includes 500 prompts across twelve risk categories, developed with expert input.
- Evaluation of seven frontier models showed failure rates between 2% and 58%, highlighting the need for improved safety measures.
- The benchmark aims to help LLM developers identify and close gaps in child safety policies.
Read more
CAREBench: A Child-Safety Risk Benchmark for Language Models
Summary
The paper introduces CAREBench, a benchmark designed to evaluate child-safety risks in language models (LMs) before they escalate into explicit harm. Unlike existing evaluations that focus primarily on child sexual abuse material, CAREBench addresses a broader spectrum of risks that can occur in interactions between LMs and children. The benchmark consists of 500 prompts categorized into twelve risk areas, including grooming, impersonation, emotional dependency, and mental health sensitivity. Developed with input from parents and clinicians, CAREBench assesses whether LMs can recognize, refuse, de-escalate, or redirect risky interactions. The evaluation of seven advanced LMs revealed failure rates ranging from 2% to 58%, indicating significant variability in how models handle different types of risks. This benchmark aims to provide LLM developers with a structured tool to identify and mitigate gaps in child safety policies, ensuring that AI systems can interact with minors responsibly.
Methodology
The authors created a structured prompt corpus validated by experts, consisting of 500 prompts that reflect various risk categories and elicitation styles. The responses from language models were evaluated based on their ability to recognize and appropriately respond to embedded child-safety risks, using annotations from parents and clinicians for grading.
Results
The evaluation of seven advanced language models revealed failure rates ranging from 2% to 58%, with distinct patterns of failure across different risk categories. This indicates that while some models perform adequately, there are significant gaps in how they handle child-safety risks.
Implications
CAREBench provides a critical tool for LLM developers to assess and improve the safety of AI systems interacting with children. By identifying specific areas of failure, developers can implement targeted interventions to enhance child safety in AI applications, ultimately contributing to more responsible AI deployment in environments frequented by minors.
Quantum Generative Diffusion Model for Real-World Time Series
Generative Models
Time Series
- Introduction of QDiffusion-TS, a quantum generative diffusion model for time series.
- Integration of quantum neural networks reduces trainable parameters significantly.
- Demonstrated superior performance in generating synthetic financial time series data.
- Augmented forecasting tasks show substantial improvements in predictive accuracy.
Read more
Quantum Generative Diffusion Model for Real-World Time Series
Summary
This paper introduces QDiffusion-TS, the first quantum generative diffusion model specifically designed for time series synthesis. The authors address the computational challenges posed by classical generative models by leveraging quantum machine learning techniques. QDiffusion-TS enhances a classical diffusion architecture by integrating quantum neural networks (QNNs) into the denoising transformer, significantly reducing the number of trainable parameters by nearly three orders of magnitude. The model is validated on financial time series data from Apple and Amazon, demonstrating its ability to generate synthetic data that closely mirrors real distributions, as evidenced by a 44% reduction in Wasserstein distance compared to classical models. Additionally, when used to augment a forecasting task, the generated data improved predictive performance by up to 71% in RMSE over a baseline that relied solely on real data. These findings suggest that quantum-enhanced architectures can not only match but often exceed the performance of classical models while maintaining a more efficient parameter count, paving the way for scalable generative modeling in data-driven applications.
Methodology
The authors developed QDiffusion-TS by replacing classical feed-forward components in a diffusion architecture with quantum neural networks, creating a hybrid quantum transformer. This approach allows for a significant reduction in the number of parameters while enhancing the model's expressiveness and efficiency.
Results
QDiffusion-TS achieved a 44% reduction in Wasserstein distance when generating synthetic time series data from Apple and Amazon compared to classical models. In downstream forecasting tasks, the model's generated data improved predictive performance by up to 71% in RMSE over a baseline model trained only on real data.
Implications
The results indicate that quantum generative models can provide a more efficient alternative to classical approaches, particularly in complex data synthesis tasks such as financial time series generation. This could lead to advancements in various applications, including finance, economics, and other fields requiring accurate time series modeling.
VGB for Masked Diffusion Model: Efficient Test-time Scaling for Reward Satisfaction and Sample Editing
Generative Models
Efficient ML
Theory
- MDM-VGB introduces a flexible sampling method that allows for arbitrary unmasking and remasking of tokens.
- The method is built on the Jerrum-Sinclair backtracking Markov chain, enhancing its applicability to structured generation tasks.
- MDM-VGB achieves quadratic complexity, outperforming traditional methods like best-of-N in terms of efficiency.
- Empirical results show significant improvements in generating valid configurations for tasks like Sudoku and QM9.
Read more
VGB for Masked Diffusion Model: Efficient Test-time Scaling for Reward Satisfaction and Sample Editing
Summary
This paper introduces MDM-VGB, a novel discrete diffusion sampler designed to enhance the performance of Masked Diffusion Models (MDM) during inference by incorporating reward-guided remasking. The authors build upon the classical Jerrum-Sinclair backtracking Markov chain to create a flexible sampling method that allows for both unmasking and remasking of tokens at arbitrary positions in a masked-state graph. This approach addresses the challenges of generating outputs that satisfy structural constraints and optimizing for downstream rewards. MDM-VGB is shown to be robust against process-verifier noise and achieves quadratic complexity, contrasting with existing methods like best-of-N, which can suffer from exponential complexity due to error accumulation. Empirical evaluations demonstrate significant performance improvements on constraint-satisfaction and scientific benchmarks, such as Sudoku and QM9, highlighting the effectiveness of the proposed method in generating high-reward configurations and repairing low-reward samples.
Methodology
The authors developed MDM-VGB by extending the classical Jerrum-Sinclair backtracking Markov chain to operate on masked-state graphs. The Markov chain allows for the revealing and remasking of tokens based on their estimated value, prioritizing configurations that lead to higher rewards. This method supports both generation from a fully masked state and editing from a low-reward fully revealed state, enabling efficient corrections of earlier mistakes.
Results
MDM-VGB and its momentum variant, MDM-VGB-Momentum, demonstrated consistent performance gains across various benchmarks, particularly in scientific design and combinatorial constraint generation tasks. The method proved to be effective in generating high-reward configurations and repairing low-reward samples, achieving quadratic convergence rates compared to exponential rates of traditional methods.
Implications
The findings suggest that MDM-VGB can significantly enhance the performance of generative models in tasks requiring structural integrity and reward optimization. This has potential applications in fields such as drug design, programming, and any domain where generating valid configurations is critical.
Continual Learning for Sequential Personalization of Small Language Models: A Stability Monitoring Analysis
NLP
Large Language Models
Efficient ML
- Introduces a checkpoint-level stability monitoring approach for SLMs during sequential personalization.
- Demonstrates that task-level performance metrics can hide detrimental adaptation effects.
- Identifies KL Divergence as an early-warning signal for model stability.
- Provides empirical analysis across multiple SLM families and tasks.
Read more
Continual Learning for Sequential Personalization of Small Language Models: A Stability Monitoring Analysis
Summary
This paper investigates the challenges of continual learning in Small Language Models (SLMs) as they adapt to user-specific data over time. The authors highlight the risk of catastrophic forgetting, where new learning can degrade performance on previously acquired tasks. They propose a checkpoint-level monitoring protocol for evaluating the stability of SLMs during sequential personalization using a lightweight adaptation method called LoRA. By saving model checkpoints after each adaptation, the authors assess performance on current tasks, previously learned tasks, and a fixed reference set. Their findings reveal that traditional task-level metrics may obscure harmful adaptation effects, and they introduce KL Divergence as a potential early-warning signal for model health. The study emphasizes the importance of reference set monitoring as a diagnostic tool for ensuring the stability of SLMs in continual learning scenarios.
Methodology
The authors operationalize the sequential personalization of SLMs as a stability monitoring problem, employing a lightweight parameter-efficient method (LoRA) for adaptation. They evaluate model performance at multiple checkpoints across tasks, creating a checkpoint-by-task evaluation matrix that tracks acquisition, retention, transfer, and stability.
Results
The study finds that monitoring stability through checkpoint evaluations reveals critical insights into model behavior that are not apparent from final accuracy metrics alone. The analysis across three SLM families and three tasks uncovers an order-invariant internal stability signature, suggesting that KL Divergence can effectively signal potential model degradation.
Implications
The findings suggest that implementing checkpoint-level monitoring can enhance the reliability and performance of SLMs in real-world applications, particularly in edge computing scenarios where continual adaptation is necessary. This research opens new avenues for developing robust monitoring tools for SLMs in continual learning contexts.
Heads, Not Backbones: Output Heads Dominate Architectures on Fat-Tailed Returns
Time Series
- Output heads significantly impact forecasting accuracy for fat-tailed financial returns.
- The Gaussian mixture density head provides substantial improvements over the single-Gaussian head, especially in volatile market conditions.
- Backbone architecture variations have minimal effects on point-prediction accuracy compared to the output head.
- The dominance of the output head is particularly pronounced at short forecasting horizons.
Read more
Heads, Not Backbones: Output Heads Dominate Architectures on Fat-Tailed Returns
Summary
This paper investigates the relative importance of backbone architectures versus output heads in deep forecasting models for fat-tailed financial returns, specifically focusing on S&P 500 monthly log-returns from 1871 to 2023. The authors compare four backbone architectures (TimesNet, DLinear, N-BEATS, iTransformer) with three types of output heads: a point head, a single-Gaussian density head, and a Gaussian mixture density head. The study employs anchored walk-forward validation to assess the models' performance using Continuous Ranked Probability Score (CRPS) and other metrics. Results indicate that the choice of output head significantly influences forecasting accuracy, with the Gaussian mixture head outperforming the single-Gaussian head, particularly in high-volatility regimes. In contrast, variations in backbone architectures yield minimal differences in performance. The findings suggest that for short-horizon forecasts of fat-tailed returns, the output head is more critical than the backbone architecture, emphasizing the need for proper scoring rules in risk management and capital allocation.
Methodology
The authors conducted a comparative analysis of 12 model variants, combining four backbone architectures with three output heads. They utilized anchored walk-forward validation on historical S&P 500 monthly log-returns data, measuring performance through CRPS, mean absolute error (MAE), coverage, and pinball loss. The study also replicated the analysis across different asset classes and time frequencies to assess generalizability.
Results
The results revealed a strict gradient in CRPS improvements when transitioning from point heads to Gaussian heads and then to mixture heads, with improvements of approximately 1.3% and 2.4%, respectively. Variations in backbone architectures resulted in less than 1.5% change in CRPS for point heads, while the density-head backbone spread reached up to 5.1%. The mixture head's value was most pronounced during high-volatility periods, confirming its effectiveness in capturing tail risk.
Implications
The findings underscore the necessity of selecting appropriate output heads for financial forecasting, particularly in risk management scenarios where tail risk is critical. The results advocate for the adoption of Gaussian mixture models in environments characterized by fat-tailed distributions, enhancing the accuracy of risk assessments and capital allocation decisions.
Layerwise Progressive Freezing: A Training Scaffold for Depth-Scalable Binary Networks
Efficient ML
- Introduction of StoMPP, a layerwise progressive freezing method for training BNNs.
- Demonstration that progression order is critical for maintaining performance in deep networks.
- Identification of activation-induced gradient blockades as a key challenge in BNN training.
- StoMPP shows significant performance gains over vanilla STE, particularly in deeper networks.
Read more
Layerwise Progressive Freezing: A Training Scaffold for Depth-Scalable Binary Networks
Summary
This paper addresses the challenges of training binary neural networks (BNNs) from scratch, particularly the accuracy degradation associated with the straight-through estimator (STE) as network depth increases. The authors introduce StoMPP (Stochastic Masked Partial Progressive Binarization), a novel training method that progressively binarizes weights and activations layer by layer, using stochastic partial masks with soft refresh. This approach allows for a fully STE-free training process that significantly outperforms vanilla STE, especially as network depth increases. The study reveals that the order of progression during training is crucial: forward layerwise progression maintains performance in deep networks, while reverse progression leads to poor outcomes. The paper also discusses the phenomenon of activation-induced gradient blockades, which hinder gradient flow in BNNs. The authors provide extensive evaluations demonstrating that StoMPP achieves substantial improvements over existing methods across various architectures, including ResNet and MobileNet, indicating its broad applicability in the field of binary neural networks.
Methodology
The authors propose StoMPP, which gradually binarizes weights and activations from input to output using a layerwise approach. The method employs stochastic masking and soft refresh to ensure that the layer currently transitioning has an unfrozen suffix, thus maintaining a gradient path to the loss. The study also includes ablation experiments in an STE-free regime to isolate the effects of the progression structure from those of surrogate gradients.
Results
StoMPP improves performance over vanilla STE by +18.0/+13.5/+3.8 on CIFAR-10/100/ImageNet for ResNet-50 BNNs. When combined with STE applied only to frozen entries, StoMPP achieves even greater gains of +27.1/+19.8/+17.7. These improvements are consistent across various architectures, including ResNet-18/34, MobileNetV2, and BERT fine-tuning.
Implications
The findings suggest that careful management of binarization during training can lead to more effective training of BNNs, making them more viable for deployment in resource-constrained environments. The insights on progression order and gradient blockades could inform future research and methodologies in binary neural network training.
TeRoR: Decoupled Temporal Rotation with Relational Circular Region for Temporal Knowledge Graph Embedding
Graph Learning
Time Series
- TeRoR introduces decoupled temporal evolution for subject and object entities, enhancing temporal information representation.
- The model employs a relation-aware circular region to effectively capture complex multi-relational interactions.
- Experimental results show significant performance improvements over existing baseline models in key evaluation metrics.
- TeRoR addresses the limitations of previous models in modeling diverse relational mappings in TKGs.
Read more
TeRoR: Decoupled Temporal Rotation with Relational Circular Region for Temporal Knowledge Graph Embedding
Summary
The paper introduces TeRoR, a novel temporal knowledge graph embedding method designed to address the limitations of existing models like TeRo in capturing complex relational mappings and temporal information. TeRoR enhances the representation of temporal knowledge graphs (TKGs) by decoupling the temporal evolution of entity embeddings, allowing for independent rotation transformations on head and tail entities in a complex vector space. This approach improves the modeling of temporal information and accommodates diverse relational characteristics by training a radius parameter that constrains the head entities within a circular region centered on the tail entity. The authors conducted systematic experiments on four distinct TKG datasets, demonstrating that TeRoR outperforms state-of-the-art models in temporal link prediction tasks, thereby validating its effectiveness in capturing intricate temporal dependencies and multi-relational interactions.
Methodology
TeRoR employs a decoupled approach to temporal evolution, allowing independent transformations for head and tail entities. It utilizes a circular region parameterized by relational characteristics to constrain the embeddings, effectively modeling complex relational mappings. The model is trained and evaluated on various TKG datasets to assess its performance in link prediction tasks.
Results
TeRoR achieved competitive performance against state-of-the-art models across four distinct TKG datasets, demonstrating stable and significant improvements in metrics such as Mean Reciprocal Rank (MRR) and Hits@K.
Implications
The advancements presented by TeRoR could enhance applications in areas requiring accurate temporal reasoning and relational understanding, such as recommendation systems, information retrieval, and cognitive reasoning tasks involving dynamic knowledge.
SP-CACW: Convergence-Aware Client Weighting for Selfish Personalized Learning
Federated Learning
Optimization
Theory
- Introduction of SP-CACW, a framework for selfish personalization in federated learning.
- The method minimizes an upper bound on the target client's convergence error, allowing for optimal client aggregation weights.
- SP-CACW can assign zero weight to peers that negatively impact convergence.
- The framework shows consistent improvements over existing methods on various datasets.
Read more
SP-CACW: Convergence-Aware Client Weighting for Selfish Personalized Learning
Summary
The paper addresses the challenges of collaborative learning in federated settings, particularly focusing on selfish personalization where a target client aims to optimize its own performance using peer gradients. The authors propose SP-CACW, a convergence-aware client-weighting framework that determines aggregation weights by minimizing an upper bound on the target client's convergence error. This method allows for the assignment of zero weight to harmful peers, effectively balancing peer bias against stochastic variance. The framework is evaluated on datasets such as MNIST, CIFAR-100, and LEAF Shakespeare, demonstrating competitive or improved performance over existing personalized and clustering-based methods. The authors provide theoretical convergence guarantees under certain assumptions, highlighting the framework's potential to enhance training dynamics for individual clients in federated learning environments.
Methodology
The authors developed a convergence-aware client weighting mechanism that explicitly derives aggregation weights based on their utility in accelerating the target client's training. The method minimizes an upper bound on the convergence rate, allowing for a principled approach to client selection in federated learning.
Results
SP-CACW was evaluated on MNIST, CIFAR-100, and LEAF Shakespeare datasets, showing that it is competitive with or improves upon strong personalized and clustering baselines. The method's theoretical guarantees indicate a convergence rate that leverages cluster size to reduce variance, enhancing training stability.
Implications
The proposed framework has significant implications for selfish federated learning scenarios, such as in inter-silo networks and public data exploitation, where clients can optimize their individual objectives while maintaining privacy and reducing negative transfer. This could lead to more effective collaboration and knowledge sharing among clients with heterogeneous data distributions.
Learning to Distributedly Estimate under Partially Known Dynamics: A Covariance-Agnostic Neural Kalman Consensus Filter
Robotics
Optimization
Theory
- Introduction of CA-NKCF, a novel distributed estimator that combines model-informed filtering with neural networks.
- Lightweight communication structure that reduces bandwidth requirements by only exchanging state priors.
- Joint optimization of NN filtering modules and consensus weights to address nonstationarity in multi-agent optimization.
- Robust performance across various environments, including linear, chaotic, and practical wireless tracking scenarios.
Read more
Learning to Distributedly Estimate under Partially Known Dynamics: A Covariance-Agnostic Neural Kalman Consensus Filter
Summary
This paper presents a novel online distributed sensing framework aimed at latent state estimation in multi-agent systems, where agents collaborate to exchange information. The proposed Covariance-Agnostic Neural Kalman Consensus Filter (CA-NKCF) integrates partial domain knowledge with deep neural network capabilities, allowing for decentralized inference without requiring precise noise statistics. The framework employs prior estimates, optimized consensus weights, and Kalman-like recursive updates to enhance estimation accuracy. Extensive experiments demonstrate that CA-NKCF outperforms traditional distributed Kalman and particle filters, as well as purely model-free deep neural networks, showcasing robustness against model misspecifications and varying noise levels. The findings indicate that CA-NKCF maintains stable performance across different communication topologies and observation conditions, making it suitable for applications in robotics and large-scale sensor networks.
Methodology
The authors developed CA-NKCF by integrating principles from multi-agent reinforcement learning with model-informed neural network filtering. The framework employs a consensus mechanism that allows agents to reach agreement on state estimates while minimizing local errors. The NN filtering modules and consensus weights are optimized using a central dataset, following the Centralized Learning, Decentralized Execution paradigm.
Results
The CA-NKCF demonstrated superior performance compared to traditional distributed Kalman filters and model-free neural networks in various experimental settings. It showed robustness to noise and model inaccuracies, maintaining consistent performance across different latent state dimensions and communication topologies.
Implications
The proposed CA-NKCF framework has significant implications for decentralized decision-making in robotics and sensor networks, enabling efficient and robust state estimation in environments where communication bandwidth is limited and system dynamics are partially known.
A Kernel Fisher Discriminant Analysis-Based Tree Ensemble Classifier: KFDA Forest
Theory
- KFDA Forest combines kernel Fisher discriminant analysis with tree-based ensemble methods to improve classification accuracy.
- The method utilizes bootstrap sampling and random variable subset division to promote diversity among classifiers.
- KFDA effectively handles nonlinear data structures, enhancing the performance of decision trees as base classifiers.
- Empirical results show that KFDA Forest outperforms traditional ensemble methods on various real datasets.
Read more
A Kernel Fisher Discriminant Analysis-Based Tree Ensemble Classifier: KFDA Forest
Summary
This paper introduces a novel ensemble classifier called the Kernel Fisher Discriminant Analysis Forest (KFDA Forest), which integrates kernel Fisher discriminant analysis (KFDA) with a tree-based ensemble method. The KFDA Forest aims to enhance classification accuracy by maximizing the distance between classes while minimizing the distance within classes, leveraging the kernel trick to handle nonlinear data structures. The methodology involves using bootstrap sampling and dividing variable sets into K subsets, on which KFDA is performed to create diverse classifiers. The authors compare KFDA Forest against existing ensemble methods, including bagging, AdaBoost, random forest, and Rotation Forest, using real datasets from UCI and KEEL repositories. The results demonstrate that KFDA Forest outperforms traditional ensemble methods, particularly in scenarios where class information is crucial for classification, thus addressing the accuracy-diversity dilemma prevalent in ensemble learning.
Methodology
The KFDA Forest employs kernel Fisher discriminant analysis for feature extraction, utilizing bootstrap sampling to create diverse subsets of data. Each subset undergoes KFDA to generate classifiers, which are then aggregated through majority voting to produce final predictions.
Results
The experimental results indicate that KFDA Forest achieves superior classification performance compared to existing methods like bagging, AdaBoost, random forest, and Rotation Forest, particularly in complex data structures where class information is significant.
Implications
The KFDA Forest has potential applications in various fields requiring robust classification methods, such as medical diagnosis, image recognition, and any domain where nonlinear relationships in data are prevalent. Its ability to enhance classification accuracy while maintaining diversity among classifiers could lead to improved decision-making processes in practical applications.
ReGuide: From Test-Time Guidance to Self-Improving Diffusion Policies
Robotics
Reinforcement Learning
Generative Models
- ReGuide transforms guided rollouts into reusable training data for self-improvement.
- Phase-Conditioned Guidance (PCG) ensures reliable corrective rollouts by focusing on phase-specific targets.
- The framework significantly outperforms existing test-time guidance methods like LPB.
- ReGuide can iteratively improve policy performance without requiring additional expert demonstrations.
Read more
ReGuide: From Test-Time Guidance to Self-Improving Diffusion Policies
Summary
The paper introduces ReGuide, a novel framework designed to enhance the performance of behavior-cloned diffusion policies, which are prone to covariate shift. Traditional methods either require expert corrections or discard guided trajectories after execution. ReGuide innovatively recycles guided rollouts as on-policy recovery data, allowing for iterative self-improvement of the policy. It employs Phase-Conditioned Guidance (PCG) to generate corrective rollouts that are phase-specific and only applied in recoverable scenarios. Successful rollouts are integrated back into the policy through two mechanisms: ReGuide-FT (fine-tuning) and ReGuide-FS (retraining from scratch). The framework demonstrates significant improvements in task success rates across various Robomimic tasks, showcasing its effectiveness in leveraging guided rollouts for continuous policy enhancement.
Methodology
ReGuide employs Phase-Conditioned Guidance (PCG) to generate corrective rollouts that are phase-specific and applicable only in recoverable scenarios. Successful rollouts are then integrated back into the policy using two methods: ReGuide-FT, which fine-tunes the existing policy, and ReGuide-FS, which retrains a new policy on the augmented dataset. This iterative process allows for continuous improvement of the policy based on the guided recovery data.
Results
ReGuide achieved a 1.3–7.7× improvement in task success rates on Robomimic tasks compared to the base policy. It outperformed the Latent Policy Barrier (LPB) in test-time-only settings, and ablation studies confirmed that the performance gains were primarily due to the use of guided recovery data rather than simply additional rollouts.
Implications
The ReGuide framework has significant implications for robotics and other fields where imitation learning is applied. By enabling policies to self-improve through guided rollouts, it reduces reliance on expert demonstrations and allows for more robust performance in dynamic environments. This approach could be extended to various applications in robotics, autonomous systems, and beyond.
Training Observable Control Policies to Expose Agent State Through Actions
Reinforcement Learning
Robotics
- Introduces an estimator that uses control policy outputs for state estimation.
- Enhances observability of control policies through reinforcement learning.
- Demonstrates improved estimator performance with minimal impact on task performance.
- Focuses on coordination in communication-limited scenarios.
Read more
Training Observable Control Policies to Expose Agent State Through Actions
Summary
This paper addresses the challenge of coordinating autonomous agents in scenarios where communication is limited or absent. The authors propose a novel approach that leverages the actions taken by agents as a source of information to estimate their internal states. By formulating the problem as a partially observable Markov Decision Process (POMDP), the authors develop observable control policies using reinforcement learning. The key innovation is the implementation of an estimator that utilizes the control policy outputs as observations, thereby enhancing the observability of the agent's state. The paper presents a simulation study focusing on an aircraft tracking problem, demonstrating that the proposed policies improve the performance of the state estimator while maintaining nominal task performance. This work lays the groundwork for more effective coordination among autonomous agents, particularly in communication-constrained environments.
Methodology
The authors formulate the problem as a POMDP and utilize reinforcement learning to train observable control policies. They implement an estimator that infers the agent's state based solely on the actions taken by the agent, rewarding the policy for improving the estimator's performance during a target tracking task. The effectiveness of the approach is analyzed using singular value decomposition of the observability matrix and Monte Carlo simulations.
Results
The simulation results indicate that the observable control policies significantly enhance the performance of the state estimator. The analysis shows that the policies trained for observability do not compromise the nominal performance of the tracking task, thereby validating the proposed approach.
Implications
This research has significant implications for the development of autonomous systems that operate in environments with limited communication capabilities. By enabling agents to infer states from observed actions, the approach can facilitate better coordination and collaboration among agents, improving safety and efficiency in various applications such as aerial surveillance, autonomous vehicles, and human-robot interaction.
Retroactive Advantage Correction: Closed-Form V-Trace Bias Correction for Delay-Aware RLHF
Reinforcement Learning
Theory
Optimization
- Introduces Retroactive Advantage Correction (RAC) to handle delayed rewards in RLHF.
- Proves that RAC can achieve unbiased corrections under specific conditions.
- Demonstrates a significant reduction in policy bias (up to 47.9×) in a tabular MDP setting.
- Integrates seamlessly with existing reinforcement learning algorithms like PPO and GRPO.
Read more
Retroactive Advantage Correction: Closed-Form V-Trace Bias Correction for Delay-Aware RLHF
Summary
This paper addresses the challenges of delayed rewards in Reinforcement Learning from Human Feedback (RLHF), particularly in production environments where synchronous reward signals are not guaranteed. The author introduces a novel method called Retroactive Advantage Correction (RAC), which allows for the correction of bias that arises when rewards are received after a delay. RAC operates by queuing slow reward signals, applying a non-negative aging kernel, and reinjecting these signals as clipped residuals into the next optimization step's advantage calculation. The paper proves that under certain conditions, the cumulative correction from RAC is unbiased when all reward mass is reinjected, and it demonstrates a linear bias in cases where some mass is not reinjected. The method is shown to reduce closed-form policy bias significantly in a tabular Markov Decision Process (MDP) setting, outperforming traditional wait-for-slow approaches while maintaining lower wall-clock costs. The integration of RAC with existing algorithms like Proximal Policy Optimization (PPO) and Generalized REINFORCE Policy Optimization (GRPO) is also discussed, highlighting its practical applicability in real-world scenarios.
Methodology
The methodology involves queuing slow reward signals that arrive after a delay, applying a non-negative kernel to age these signals, and reinjecting them into the advantage calculation of the next optimization step. The approach leverages clipped importance sampling ratios to ensure unbiasedness and integrates with existing reinforcement learning frameworks.
Results
In a proof-of-concept using a tabular MDP, RAC achieved a reduction in closed-form policy bias by up to 47.9× in a configuration with two slow channels. The method was validated through multiple machine-precision checks, confirming the theoretical claims regarding bias correction and the equivalence to V-trace under certain conditions.
Implications
The findings suggest that RAC can significantly improve the performance of RLHF systems in real-world applications where reward signals are often delayed. This could lead to more efficient training processes and better overall policy performance in various reinforcement learning tasks.
The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patching
Interpretability
Theory
Large Language Models
- Natural Indirect Effect (NIE) includes interaction effects (INT) that can misrepresent component importance.
- INT scales with the difference between clean and patched activations and is negligible in locally affine models.
- The presence of INT explains known failures in activation patching, particularly in multi-component systems.
- Ranking components solely by PIE (pure indirect effect) can lead to significant misinterpretations of their importance.
Read more
The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patching
Summary
This paper investigates the limitations of activation patching, a method used in mechanistic interpretability to attribute causal responsibility in neural networks. The authors derive the natural indirect effect (NIE) from causal mediation analysis and reveal that it not only captures the causal effect through specific components but also includes interaction effects (INT) that depend on the states of other components. They demonstrate that these INT can lead to misleading conclusions about component importance, particularly in complex models like transformers. The study highlights that INT is not merely a nuisance but a crucial diagnostic tool for interpretability, indicating when causal conclusions are prompt-dependent. The authors provide theoretical insights and empirical evidence showing that the presence of INT can significantly alter the ranking of component importance, and they discuss the implications of these findings for the reliability of activation patching as a causal attribution method.
Methodology
The authors re-derive the activation patching estimator from causal mediation analysis and analyze the interaction effects in transformer models. They conduct theoretical proofs and empirical evaluations using the GPT-2 IOI circuit to illustrate the impact of INT on component importance rankings.
Results
The study finds that the NIE and PIE yield different rankings of component importance, with correlation coefficients as low as 0.51. INT is shown to vary significantly based on the state of other components and is responsible for known pathologies in activation patching. The authors demonstrate that removing INT leads to its own ranking failures, particularly for context-dependent components.
Implications
These findings suggest that practitioners should account for interaction effects when using activation patching for causal attribution in neural networks. Understanding INT can improve the reliability of interpretability studies and guide more effective component evaluations in complex models.
Statistically Indistinguishable, Operationally Distinct: A Formal Barrier for Tabular Foundation Models
Theory
- Tabular foundation models cannot effectively reason about data without access to operational rules.
- The Operational Turing Test (OTT) establishes a formal framework to evaluate this limitation.
- Values-only classifiers are statistically indistinguishable under certain conditions, leading to a Bayes error of at least 0.49.
- Exposing relational value consistency improves model performance but does not eliminate the need for operational context.
Read more
Statistically Indistinguishable, Operationally Distinct: A Formal Barrier for Tabular Foundation Models
Summary
This paper investigates the limitations of tabular foundation models in reasoning about data generated by running systems without access to the operational rules governing them. The authors introduce the Operational Turing Test (OTT), which constructs pairs of legal and rule-violating database states that are statistically indistinguishable based on their column-value marginals. They demonstrate that any values-only classifier will have a Bayes error of at least 0.49 when distinguishing between these states. The study evaluates three values-only classifiers (XGBoost, TabICL, TabPFN) which all achieve an accuracy of 0.50, confirming the theoretical bound. The authors find that while exposing relational value consistency improves performance, it does not fully bridge the gap; only classifiers utilizing rule-derived audits achieve perfect accuracy. Furthermore, even advanced large language models (LLMs) struggle to classify legal states correctly when provided with schema and operational context. The findings highlight that the barrier to effective classification lies in the identifiability of operational rules rather than the capacity of the models, suggesting that richer features or larger models cannot overcome this limitation without operational grounding.
Methodology
The authors develop the Operational Turing Test (OTT) to assess the ability of classifiers to distinguish between legal and illegal database states based on their statistical properties. They employ Le Cam's lemma to establish a Bayes error bound for values-only classifiers and conduct empirical tests using three baseline classifiers and rule-derived audits to evaluate classification accuracy.
Results
The study finds that values-only classifiers achieve an accuracy of 0.50, consistent with the theoretical Bayes error bound of 0.49. While exposing relational value consistency improves accuracy, it does not reach the optimal classification level. Only classifiers that incorporate operational rules achieve perfect accuracy. Additionally, frontier LLMs struggle to classify legal states correctly, even when provided with extensive operational context.
Implications
The findings suggest that for tabular foundation models to be effective in real-world applications, they must incorporate operational grounding and access to the rules governing data generation. This has significant implications for the design of future machine learning systems, particularly in enterprise contexts where operational logic is critical.
Position: RL Researchers Need to Distinguish Between Solving Simulators and Using Simulators as a Proxy
Reinforcement Learning
- Distinction between solving simulators and using them as proxies for real-world learning is crucial.
- Misunderstanding these goals can lead to misleading conclusions in RL research.
- Different algorithms and evaluation metrics are appropriate for each use case.
- The paper calls for the RL community to clarify their use of simulators in research.
Read more
Position: RL Researchers Need to Distinguish Between Solving Simulators and Using Simulators as a Proxy
Summary
This paper argues that reinforcement learning (RL) researchers must differentiate between two distinct use cases of simulators: solving simulators and using simulators as a proxy for learning in deployment settings. The authors highlight how the focus on achieving high performance in simulators can lead to a misunderstanding of the true goals of RL research, which should be centered around learning in real-world environments. They discuss the constraints and appropriate algorithms for each use case, emphasizing that treating simulators merely as benchmarks can result in misleading conclusions and hinder the development of algorithms that generalize well to deployment scenarios. The paper provides examples and simple experiments to illustrate the consequences of conflating these two goals and calls for clearer distinctions in empirical practices within the RL community to foster better research outcomes.
Methodology
The authors discuss the theoretical framework surrounding the use of simulators in RL, providing examples and conducting simple experiments to illustrate the differences between solving simulators and using them as proxies for deployment learning. They analyze existing literature and practices in the field to highlight common pitfalls and misconceptions.
Results
The paper does not present empirical results in the traditional sense but rather emphasizes the conceptual framework and implications of distinguishing between the two use cases. It showcases how failing to make this distinction can lead to suboptimal algorithm development and evaluation.
Implications
Clarifying the distinction between the two use cases can lead to more robust RL algorithms that are better suited for real-world applications. It encourages researchers to adopt practices that prioritize learning in deployment settings, potentially improving the effectiveness of RL solutions in practical scenarios.
Estimation--Prediction Tradeoff in Causal Probabilistic Temporal Graphs
Graph Learning
Theory
Time Series
- Introduces an estimation–prediction tradeoff in binary logistic models for temporal link prediction.
- Proposes a probabilistic causal framework for generating temporal graphs with known causal structures.
- Derives the Cramér–Rao bound to analyze the relationship between parameter estimation error and predictive performance.
- Highlights that high predictive accuracy does not necessarily indicate successful learning of causal mechanisms.
Read more
Estimation--Prediction Tradeoff in Causal Probabilistic Temporal Graphs
Summary
This paper addresses the challenges in temporal link prediction within probabilistic temporal graphs, particularly the conflation of model error with irreducible uncertainty. The author characterizes an inherent estimation–prediction tradeoff in binary logistic models, where regimes maximizing Fisher information also exhibit higher entropy, complicating individual predictions despite perfect parameter recovery. A novel probabilistic causal framework is proposed for generating temporal graphs with transient edges and known causal structures, allowing for a joint evaluation of temporal link prediction and causal parameter recovery. The paper derives the Cramér–Rao bound for the proposed binary logistic parametrization and validates the tradeoff between parameter estimation error and predictive loss. The findings suggest that predictive accuracy alone does not guarantee that a model has learned the underlying causal mechanisms, advocating for benchmarks that differentiate reducible model error from intrinsic process uncertainty.
Methodology
The paper employs a probabilistic causal framework to generate temporal graphs, allowing for the evaluation of temporal link prediction alongside causal parameter recovery. It utilizes information theory to derive the Cramér–Rao bound and analyze the estimation–prediction tradeoff in binary logistic models.
Results
The analysis reveals that regimes that enhance parameter estimation accuracy also lead to increased predictive difficulty due to higher entropy. Empirical results demonstrate that predictive performance does not always correlate with successful causal discovery, emphasizing the need for refined evaluation metrics.
Implications
The findings have significant implications for the fields of causal inference and temporal graph analysis, suggesting that models should be evaluated not only on predictive accuracy but also on their ability to recover underlying causal structures. This could lead to improved methodologies in various applications, including social network analysis and dynamic systems modeling.
COCOLogic-V2: Identifying Logical Inconsistencies via Truly Hard-Negatives
Computer Vision
Interpretability
- Introduction of COCOLogic-V2, a dataset for visual inductive reasoning on real-world images.
- Categorization of samples into positive variants, near-boundary, and far-from-boundary negatives for detailed model evaluation.
- Reformulation of the task as multilabel classification to reduce class imbalance and logical shortcuts.
- Demonstration that current interpretable models struggle with near-boundary samples, highlighting the need for improved reasoning capabilities.
Read more
COCOLogic-V2: Identifying Logical Inconsistencies via Truly Hard-Negatives
Summary
The paper introduces COCOLogic-V2, an object-centric dataset designed for visual inductive reasoning on real-world images, expanding upon the limitations of the original COCOLogic dataset. COCOLogic-V2 addresses the challenges of evaluating interpretable models like concept bottleneck models (CBMs) and neuro-symbolic approaches, which have been primarily tested on simpler tasks. The dataset categorizes samples into positive variants, near-boundary (NB) negatives, and far-from-boundary (FB) negatives, enabling a nuanced assessment of model accountability. The authors reformulate the task as multilabel classification to mitigate class imbalance and shortcuts that hinder logical reasoning. Evaluations reveal that while models can effectively distinguish positive and FB samples, they struggle with NB samples, indicating a reliance on statistical patterns rather than genuine rule comprehension. The paper also presents COCOLogic-V2-FS, a smaller version for few-shot learning scenarios. Overall, COCOLogic-V2 serves as a foundational tool for advancing visual inductive reasoning methods in complex real-world contexts.
Methodology
The authors developed COCOLogic-V2 by expanding the logical scope of the original COCOLogic dataset to include complex first-order logical operations. They categorized samples into distinct groups to facilitate fine-grained model evaluation and diagnosis. The dataset was evaluated using various model families, including CBMs and black-box baselines, to assess their performance on both COCOLogic-V2 and its few-shot variant.
Results
The evaluations indicated that models performed well in separating positive samples from FB negatives but consistently failed with NB negatives. This suggests that models are leveraging statistical correlations rather than understanding the underlying logical rules. Additionally, challenges such as perceptual noise and large search spaces were noted in few-shot learning contexts.
Implications
COCOLogic-V2 provides a robust framework for testing and improving visual inductive reasoning in machine learning models, particularly in high-stakes domains where accountability and transparency are critical. The findings underscore the need for advancements in model interpretability and reasoning capabilities.
TA-SparseMG: Trend-Aware Sparse Forecasting via Multi-Scale Gating for Long-Term Time Series
Time Series
- Introduction of TA-SparseMG, a lightweight model for long-term time series forecasting.
- Incorporation of a trend-aware normalization module to mitigate distribution shifts.
- Development of a gated denoising module to reduce high-frequency noise interference.
- Utilization of a multiscale gated attention mechanism to enhance prediction adaptability.
Read more
TA-SparseMG: Trend-Aware Sparse Forecasting via Multi-Scale Gating for Long-Term Time Series
Summary
The paper introduces TA-SparseMG, a lightweight model designed for long-term time series forecasting (LTSF) that addresses challenges such as statistical nonstationarity and high-frequency disturbances. It builds upon the SparseTSF framework and incorporates three innovative modules: a trend-aware reversible instance normalization (TA-RevIN) module that adjusts for distribution shifts, a scale-adaptive gated denoising module that smooths features and suppresses noise, and a multiscale gated attention MLP predictor that enhances the model's representational capacity through conditional gating. The proposed model aims to maintain parameter efficiency while improving forecasting accuracy and robustness. Extensive experiments across multiple LTSF benchmarks demonstrate that TA-SparseMG outperforms existing models, with ablation studies confirming the contribution of each module to the overall performance. This work highlights the importance of addressing distribution adaptation, feature denoising, and dynamic prediction in LTSF tasks.
Methodology
The methodology involves a three-module architecture: (1) TA-RevIN for capturing intra-series statistical dynamics, (2) a scale-adaptive gated denoising module for feature smoothing and noise suppression, and (3) a multiscale gated attention MLP for adaptive feature transformation. The model leverages a sparse cross-period modeling framework to ensure parameter efficiency while addressing the complexities of long-term forecasting.
Results
TA-SparseMG consistently achieves superior performance across six mainstream LTSF benchmarks, demonstrating improved forecasting accuracy and stability compared to existing models. Ablation studies validate the effectiveness of each module, confirming their individual contributions to enhancing distribution adaptation and robustness.
Implications
The findings suggest that TA-SparseMG can be effectively applied in various domains requiring long-term forecasting, such as energy demand prediction, traffic flow analysis, and meteorological forecasting. The model's lightweight nature makes it suitable for real-time applications where computational efficiency is critical.
IG-Lens: Exact Additive Probability Attribution Across Transformer Layers via Telescoping Integrated Gradients
NLP
Large Language Models
Interpretability
- IG-Lens provides the first exact additive decomposition of probability across transformer layers.
- The method integrates the softmax function within the attribution process, avoiding biases present in previous methods.
- A prediction-aware estimator enhances the accuracy of layer-wise attributions by considering observed changes in target probability.
- The implementation is efficient, allowing for full token-by-layer mapping without backward calls.
Read more
IG-Lens: Exact Additive Probability Attribution Across Transformer Layers via Telescoping Integrated Gradients
Summary
The paper introduces IG-Lens, a novel method for exact additive probability attribution in decoder-only transformers. It addresses the challenge of determining the specific layers in a transformer where the probability of a predicted token is produced, a question inadequately answered by existing methods. Previous tools, such as the logit lens and Direct Logit Attribution, either provide biased estimates or fail to account for the nonlinearity introduced by the softmax function. IG-Lens employs a telescoping application of Integrated Gradients along a single path through the hidden states, crediting each segment to the layer it terminates at. This approach allows for a layer-wise attribution that sums to the total change in prediction probability, integrating the softmax function rather than linearizing it. The method includes a prediction-aware estimator that reweights integration steps based on observed changes in target probability, ensuring completeness and accuracy. The paper also presents a single-pass batched implementation that computes the full token-by-layer map efficiently, without requiring backward calls. Overall, IG-Lens provides a significant advancement in understanding how transformer models make predictions, with implications for targeted interventions and model interpretability.
Methodology
IG-Lens utilizes a telescoping application of Integrated Gradients along a path through the hidden states of a transformer model. It credits each segment of the path to the layer at which it ends, integrating the softmax function rather than linearizing it. A prediction-aware estimator is employed to weight integration steps based on observed changes in target probability, ensuring completeness and accuracy.
Results
The method achieves an exact additive decomposition of the probability changes across layers, demonstrating completeness to floating point precision. The implementation allows for efficient computation of the full token-by-layer map, significantly improving upon existing methods that provide only approximate attributions.
Implications
IG-Lens has potential applications in enhancing the interpretability of transformer models, enabling researchers and practitioners to understand the decision-making processes of these models better. It also opens avenues for targeted interventions, such as activation patching, by identifying the layers responsible for specific predictions.
Disentangling Continuous-Time Latent Dynamics: Identifiability of Latent SDEs via Diffusion Shifts
Time Series
Theory
- Identifiability of continuous-time latent SDEs is achieved through diffusion shifts.
- Two diagonal diffusion regimes with distinct variance ratios can anchor the latent coordinate system.
- The proposed method does not require sparsity assumptions on the drift.
- A practical two-stage estimator is developed for latent disentanglement and graph recovery.
Read more
Disentangling Continuous-Time Latent Dynamics: Identifiability of Latent SDEs via Diffusion Shifts
Summary
This paper addresses the challenge of identifiability in continuous-time latent stochastic differential equation (SDE) models, a gap in causal representation learning (CRL) that has been well-explored in discrete-time settings. The authors investigate additive-noise latent SDEs observed through an unknown nonlinear diffeomorphism, focusing on environments with shared drift but differing diffusion covariance. They demonstrate that two diagonal diffusion regimes with distinct coordinate-wise variance ratios can identify the latent coordinates up to permutation and scaling, without requiring sparsity assumptions on the drift. The results are first established for linear Ornstein–Uhlenbeck systems and then extended to general additive-noise latent SDEs. The paper introduces a two-stage estimator for latent disentanglement and optional graph recovery, validated through synthetic experiments and applied to real sensor data from the Hardanger Bridge monitoring study.
Methodology
The authors formulate the CRL problem in continuous time and identify diffusion shifts as a source of side information. They prove identifiability for linear Ornstein–Uhlenbeck systems and general additive-noise latent SDEs. A two-stage estimator is proposed, first performing diffusion-based disentanglement followed by sparse drift regression for graph recovery.
Results
The study establishes that the latent coordinate system can be identified up to permutation and scaling, and the instantaneous drift-Jacobian causal graph is identifiable under mild smoothness conditions. Synthetic experiments confirm the predicted identifiability boundary, and the method is successfully applied to real sensor data.
Implications
This work has significant implications for fields that rely on understanding latent dynamics in continuous time, such as climate science, biology, and healthcare. The ability to disentangle latent variables from nonlinear observations can enhance forecasting and decision-making processes.
Optimizer Memory Makes Shuffle Order a First-Order Source of Fine-Tuning Noise
Optimization
Theory
- Fixed-clock optimizers like AdamW introduce first-order noise due to their memory-dependent state.
- Reordering data in fine-tuning can significantly alter outcomes, contrary to traditional memoryless assumptions.
- The paper presents a fit-free method to quantify the noise produced by shuffle order.
- Local order-noise scales can be used to size shuffle-seed comparisons effectively.
Read more
Optimizer Memory Makes Shuffle Order a First-Order Source of Fine-Tuning Noise
Summary
This paper investigates the impact of shuffle order on fine-tuning noise in machine learning optimizers, particularly focusing on fixed-clock optimizers like AdamW. The author argues that traditional analyses underestimate the noise introduced by shuffle order, as they often assume a memoryless optimizer. In contrast, fixed-clock optimizers maintain a state that evolves with each step, leading to position-dependent weights for gradients. This results in a first-order noise channel that can significantly affect the outcomes of fine-tuning, potentially flipping results in close A/B comparisons. The paper derives a fit-free method to quantify this noise and demonstrates that the order of data can influence the final model performance more than previously thought. The findings highlight the necessity of considering optimizer memory when evaluating fine-tuning results, especially in the context of adaptive optimizers. The study also provides a framework for measuring order-noise error bars and positional attribution weights, which can guide fine-tuning comparisons.
Methodology
The author employs a theoretical framework to analyze the behavior of fixed-clock optimizers in the context of fine-tuning. By isolating the effects of optimizer memory and deriving mathematical expressions for noise quantification, the study utilizes a lifted-state expansion approach to assess the impact of shuffle order on model performance. Empirical evaluations are conducted using various optimizers, including AdamW and fixed-β momentum, to validate the theoretical predictions.
Results
The study finds that fixed-clock optimizers exhibit a first-order noise channel, with order-variance slopes of 1.83 for AdamW, 2.00 for fixed-β momentum, and 4.00 for SGD. The results indicate that the noise introduced by shuffle order can be substantial enough to influence fine-tuning outcomes significantly. Additionally, a closed-form asymptotic order-variance floor is derived for AdamW with a frozen preconditioner, providing a robust method for measuring local potentials and noise.
Implications
The findings suggest that practitioners should account for the effects of shuffle order and optimizer memory when conducting fine-tuning experiments. This could lead to more reliable model evaluations and improved methodologies for hyperparameter tuning. Furthermore, the proposed framework for measuring order-noise could enhance the robustness of fine-tuning comparisons across different models and datasets.
Graph Dimensionality Reduction for Contextual Bandits: Structure-Specific Regret Bounds under Approximate Smoothness and Noisy Eigenspaces
Graph Learning
Theory
Reinforcement Learning
- GraphDR-LinUCB achieves eO(k√T) regret, improving efficiency by reducing dimension dependence from d to k.
- The method incorporates a structure-specific residual analysis that mitigates worst-case penalties associated with high-frequency rewards.
- A novel spectral selection rule is introduced to predict the best reducer without needing fitted thresholds.
- Empirical results demonstrate significant reductions in cumulative regret across various datasets.
Read more
Graph Dimensionality Reduction for Contextual Bandits: Structure-Specific Regret Bounds under Approximate Smoothness and Noisy Eigenspaces
Summary
This paper addresses the challenge of contextual bandits with graph-structured arms, which are prevalent in applications like recommendation systems and social advertising. Traditional dimensionality reduction techniques overlook the graph structure, leading to inflated exploration costs. The authors propose GraphDR-LinUCB, a method that projects arm features onto a low-frequency spectral subspace of the graph and applies linear UCB in this reduced k-dimensional space. They establish the first eO(k√T) regret bound for this approach, significantly reducing the dimension dependence from d to k. The theoretical framework includes a perturbation argument that accommodates noisy graphs and quantifies the penalty for reward-smoothness mismatch. The authors demonstrate that the high-frequency reward component's cost is contingent on its impact along the played path rather than its total energy. A spectral comparison method is introduced to predict the effectiveness of different reducers on datasets. Empirical evaluations across synthetic benchmarks and six real datasets show that GraphDR-LinUCB reduces cumulative regret by 15 times compared to full-dimensional LinUCB and outperforms other graph-aware methods in five out of six cases, with the exception occurring when the graph's spectral subspace misaligns with the reward.
Methodology
The authors develop GraphDR-LinUCB by projecting arm feature vectors onto the bottom-k Laplacian eigenspace and applying the LinUCB algorithm in this reduced space. They prove regret bounds using exact-subspace and Davis–Kahan arguments, and conduct empirical evaluations with graph-shuffle controls to validate their approach.
Results
GraphDR-LinUCB reduces cumulative regret by 15 times compared to full-dimensional LinUCB and outperforms competing graph-aware methods in five out of six datasets tested. The method's theoretical foundations provide robust regret bounds under both exact and approximate smoothness conditions.
Implications
This work has significant implications for improving the efficiency of contextual bandit algorithms in graph-structured environments, potentially enhancing applications in recommendation systems, social networks, and other domains where graph relationships influence reward structures.
Convergence of Continual Learning in Homogeneous Deep Networks
Theory
- Weakly-regularized continual learning in homogeneous DNNs performs sequential margin projections.
- Global convergence is not guaranteed in homogeneous models, contrasting with linear models.
- Local convergence can be achieved through nonconvex projection theory.
- The analysis framework is extended from classification to regression, unifying the approach.
Read more
Convergence of Continual Learning in Homogeneous Deep Networks
Summary
This paper addresses the theoretical foundations of continual learning in homogeneous deep neural networks (DNNs), focusing on weakly regularized continual classification. The authors establish that continual learning in these models can be characterized as sequential projections onto task margin sets, a significant generalization from previous analyses limited to linear models or stationary settings. They demonstrate that while global convergence is generally not achievable, local linear convergence can be guaranteed under specific conditions using nonconvex projection theory. The analysis is extended to continual regression, providing a unified framework for understanding the dynamics of homogeneous models. The findings highlight the complexities of continual learning in deep networks, emphasizing the need for tailored theoretical tools to capture the nonlinear dynamics inherent in these architectures.
Methodology
The authors utilize a theoretical framework based on weak isotropic regularization and nonconvex projection theory to analyze the behavior of homogeneous DNNs during continual learning. They establish equivalences between continual classification and sequential margin-separating projection algorithms, allowing for a deeper understanding of convergence properties in these models.
Results
The paper shows that while global convergence fails for homogeneous DNNs, local linear convergence is achievable under random and cyclic task sequences. The results indicate that the projection sets in these models are not necessarily convex, leading to qualitative differences in learning dynamics compared to linear models. The extension to continual regression further solidifies the framework's applicability.
Implications
The findings have significant implications for the design of continual learning algorithms in deep learning, suggesting that practitioners should consider the nonconvex nature of projection sets in their models. This understanding can lead to improved strategies for mitigating catastrophic forgetting and enhancing model performance in dynamic environments.
From Failure Taxonomy to Intervention: A Diagnostic Methodology for Industry-Scale AVLM in Video and Live-Streaming Platform Moderation
Multimodal
Audio & Speech
- Introduction of a systematic diagnostic methodology for AVLM development.
- Categorization of model failures into interpretable taxonomies linked to actionable interventions.
- Focus on integrating audio-visual context to improve moderation accuracy.
- Development of an AVLM that supports diverse content across multiple regions.
Read more
From Failure Taxonomy to Intervention: A Diagnostic Methodology for Industry-Scale AVLM in Video and Live-Streaming Platform Moderation
Summary
This paper addresses the challenges of moderating user-generated content on video and live-streaming platforms, particularly the limitations of existing pretrained models and APIs in adapting to platform-specific requirements. The authors propose a diagnostic methodology for developing Audio-Visual-Language Models (AVLMs) that categorizes model failures into a taxonomy of observable signatures, linking each failure type to specific intervention strategies. This systematic approach aims to enhance the reliability and effectiveness of AVLMs in handling diverse and ambiguous content. The methodology is instantiated in the development lifecycle of an AVLM for a large-scale platform, which supports over 100 regions and is designed to manage the complexities of global user-generated content. The paper emphasizes the importance of targeted interventions to address deployment failures, moving away from heuristic trial-and-error methods that often obscure the underlying causes of issues. By integrating audio, visual, and contextual data, the proposed AVLM aims to improve moderation accuracy and reduce both under-enforcement and over-enforcement of content policies.
Methodology
The authors developed a diagnostic framework that maps observable model failures into a taxonomy and links these failures to specific intervention strategies. This methodology was applied throughout the development and alignment lifecycle of an AVLM, which integrates audio, visual, and contextual data for effective moderation.
Results
The resulting AVLM system demonstrated improved moderation capabilities, effectively managing over 100 regions and addressing the challenges posed by noisy and ambiguous content. The methodology facilitated targeted interventions that enhanced model reliability and reduced common moderation errors.
Implications
The proposed methodology has significant implications for the development of AI moderation systems in various industries, particularly in enhancing the accuracy and reliability of content moderation on platforms with large volumes of user-generated content. It offers a structured approach to diagnosing and addressing model failures, which can lead to better alignment with evolving content policies.
PerturbCellRL: Verifier-Guided Reinforcement Learning for Single-Cell Perturbation Prediction
Reinforcement Learning
Generative Models
- PerturbCellRL incorporates biological verifiers as reward functions to ensure individual cell responses are biologically plausible.
- The framework improves upon existing generative models by aligning predictions with biological consistency metrics.
- PerturbCellRL demonstrates superior performance on multiple benchmarks while maintaining competitive population-level metrics.
- The methodology allows for better selection of predictions through a best-of-N approach using pathway activity verifiers.
Read more
PerturbCellRL: Verifier-Guided Reinforcement Learning for Single-Cell Perturbation Prediction
Summary
The paper introduces PerturbCellRL, a novel reinforcement learning framework designed to enhance single-cell perturbation prediction by ensuring biological consistency in generated cell responses. Traditional generative models often fail to validate individual cell responses against biological realities, focusing instead on population-level predictions. PerturbCellRL addresses this limitation by employing a suite of biological verifiers that serve as reward functions during the post-training phase of a pretrained single-cell transcriptomic generator. The verifiers assess generated cells based on four criteria: Pearson top-k similarity, RMSE top-k proximity, DE Spearman, and Pathway activity, which collectively evaluate target alignment and biological plausibility. The authors demonstrate that PerturbCellRL outperforms the baseline flow-matching generator on various genetic and chemical perturbation benchmarks, achieving improved scores on reward-aligned metrics without sacrificing population-level performance. This approach not only enhances the reliability of single-cell predictions but also supports the broader goals of drug discovery and personalized medicine by reducing the need for costly wet-lab experiments.
Methodology
PerturbCellRL utilizes a pretrained flow-matching generator and employs reinforcement learning to post-train the model using a suite of biological verifiers as reward functions. The model generates multiple candidate cell responses, which are scored based on the defined rewards, and updates the generator to favor high-reward outputs. At inference, the pathway activity verifier is used for selecting the most biologically plausible predictions.
Results
The evaluation of PerturbCellRL on various genetic and chemical perturbation benchmarks shows significant improvements over the baseline flow-matching generator in terms of reward-aligned metrics and held-out evaluation metrics. The best-of-N selection process further enhances biological consistency, indicating that the integration of verifier-guided scaling yields better predictions without compromising distributional quality.
Implications
The findings suggest that incorporating biological verifiers into generative models can significantly enhance the reliability of single-cell perturbation predictions, which is crucial for applications in drug discovery and personalized medicine. This approach may reduce the reliance on expensive wet-lab experiments by providing more accurate in silico predictions.
BaRA: Bayesian Adaptive Rank Allocation for Parameter-Efficient Fine-Tuning
NLP
Large Language Models
Efficient ML
- BaRA introduces a dynamic, context-dependent rank allocation mechanism for fine-tuning large language models.
- The framework employs a hierarchical Bayesian structure to model both data and model uncertainty.
- BaRA improves predictive performance and uncertainty calibration compared to traditional LoRA and Bayesian LoRA methods.
- A complexity-theoretic analysis supports the effectiveness of BaRA in reducing model complexity.
Read more
BaRA: Bayesian Adaptive Rank Allocation for Parameter-Efficient Fine-Tuning
Summary
The paper introduces BaRA, a Bayesian Adaptive Rank Allocation framework designed to enhance parameter-efficient fine-tuning (PEFT) of large language models (LLMs). Traditional low-rank adaptation (LoRA) methods, while efficient, often suffer from limitations in representational flexibility and uncertainty calibration, particularly in low-data scenarios. Existing Bayesian LoRA variants improve uncertainty estimation but typically use fixed ranks, which do not account for the varying adaptation needs across different contexts. BaRA addresses this by dynamically allocating adaptation capacity through a hierarchical Bayesian structure that utilizes context-aware local latent variables and layer-specific global latent variables. This allows for instance-wise variation in effective rank, improving both predictive performance and uncertainty calibration. The authors also provide a complexity-theoretic generalization analysis, demonstrating that BaRA reduces the effective hypothesis complexity while maintaining expressiveness. Extensive experiments on various natural language benchmarks show that BaRA outperforms standard LoRA and existing Bayesian LoRA methods in terms of predictive performance, robustness, and uncertainty calibration.
Methodology
BaRA employs a hierarchical Bayesian framework that combines context-aware local latent variables for adaptive rank allocation and layer-specific global latent variables to promote structured sparsity. It utilizes amortized variational inference techniques to ensure scalability to large language models while maintaining Bayesian uncertainty over adaptation parameters.
Results
The experimental results demonstrate that BaRA consistently enhances predictive performance, robustness, and uncertainty calibration across various natural language processing benchmarks, outperforming both standard LoRA and existing Bayesian LoRA variants.
Implications
BaRA's approach to dynamic rank allocation could significantly improve the efficiency and effectiveness of fine-tuning large language models, making it applicable in scenarios with limited data and enhancing the reliability of model predictions.
Prototype Latent World Model Replay for Class-Incremental Learning
Computer Vision
- Introduces a memory-free framework for class-incremental learning that uses latent state distributions instead of raw images.
- Utilizes a frozen pretrained encoder to maintain a stable latent space for old-class representations.
- Implements a prototype-based class memory with multiple latent prototypes and variances.
- Achieves significant accuracy improvements on Split CIFAR-100 compared to traditional fine-tuning methods.
Read more
Prototype Latent World Model Replay for Class-Incremental Learning
Summary
The paper addresses the challenge of class-incremental learning (CIL), where models must learn new classes while retaining knowledge of old ones without access to previous samples. The authors propose a novel framework called Prototype Latent World Model Replay, which represents old classes as distributions over stable latent states rather than storing raw images. Utilizing a frozen ImageNet-pretrained encoder, the method maps images into a latent state space where each class is summarized by multiple prototype-centered distributions. When new classes are introduced, the model samples old latent states from these distributions and trains a lightweight adapter and classifier using both sampled old states and new-class features. A supervised contrastive term is incorporated to enhance intra-class compactness and separate old and new classes. The proposed method demonstrates significant improvements over traditional fine-tuning approaches on the Split CIFAR-100 dataset, achieving higher accuracy without the need for storing exemplars. The results indicate that stable latent-state replay is crucial for performance gains, and the contrastive separation further refines the model's ability to distinguish between old and new classes.
Methodology
The methodology involves mapping images into a stable latent state space using a frozen pretrained encoder. Each class is represented by a mixture of prototype-centered distributions. When new classes are introduced, the model samples latent states from these distributions and combines them with real new-class features to train a lightweight adapter and classifier. A supervised contrastive term is added to promote intra-class compactness and separate old and new classes.
Results
The proposed Ours-LWM+Con model significantly improves accuracy on the Split CIFAR-100 dataset, raising LastAcc from 4.55% to 31.64% for Inc5, from 9.06% to 37.06% for Inc10, and from 16.96% to 43.10% for Inc20. The average accuracy also increases to 45.86%, 52.19%, and 56.18% for Inc5, Inc10, and Inc20, respectively. Ablation studies indicate that stable latent-state replay is the primary source of performance improvement.
Implications
The findings suggest that the proposed framework can effectively mitigate catastrophic forgetting in class-incremental learning scenarios, making it applicable in real-world settings where models need to adapt to new classes over time without retraining on old data. This approach could be beneficial in various domains such as robotics, autonomous systems, and any application requiring continuous learning.
Beyond Sparse Supervision: Diffusion-Guided Learning for Few-Shot Graph Fraud Detection
Graph Learning
- Introduces ADC-GNN, a novel framework for few-shot graph fraud detection.
- Utilizes diffusion-guided feature augmentation to enhance representation stability.
- Combines contrastive learning with multi-hop spectral attention for improved anomaly detection.
- Demonstrates consistent performance improvements on public benchmarks and a real-world dataset.
Read more
Beyond Sparse Supervision: Diffusion-Guided Learning for Few-Shot Graph Fraud Detection
Summary
This paper addresses the challenges of detecting fraud in graph-based transaction systems, particularly in scenarios with sparse and imbalanced labels. The authors propose a novel framework called ADC-GNN (Attention-guided Diffusion-Contrastive Graph Neural Network), which integrates diffusion-guided feature augmentation, contrastive representation learning, and multi-hop spectral attention. The framework aims to enhance the detection of fraudulent activities by overcoming issues such as representation dilution and the dominance of benign accounts in training data. The diffusion component serves as a feature-space denoising mechanism, creating noise-perturbed views of node features to stabilize representations through contrastive learning. The spectral attention module emphasizes relevant cues at different hop and relation levels, improving the model's ability to discern anomalies. The effectiveness of ADC-GNN is evaluated on three public benchmarks and a proprietary telecom dataset, demonstrating significant improvements over existing methods under few-shot learning conditions. The paper also includes extensive analyses on various aspects of the model's performance, including stability and efficiency.
Methodology
The ADC-GNN framework employs a diffusion-guided feature augmentation approach to create noise-perturbed views of node features, which are then stabilized using contrastive learning. Additionally, it incorporates a multi-hop spectral attention mechanism that adaptively highlights fraud-relevant features across different levels of the graph.
Results
ADC-GNN achieves significant performance improvements over baseline models in detecting fraud in sparse and imbalanced datasets. The model consistently outperforms four recent graph anomaly detection methods across three public benchmarks and a proprietary dataset, particularly under a 1% training set scenario.
Implications
The proposed framework has the potential to enhance fraud detection in various financial and transaction systems, particularly in environments where labeled data is scarce. Its ability to effectively utilize limited supervision can lead to more robust and reliable fraud detection mechanisms.