AI-generated summaries
Today's ML research,
without the noise.
Daily summaries of the latest machine learning papers from arXiv, processed every 8 hours.
24
Papers today
8h
Update frequency
7
Days of history
Frozen Cores Need Task Signal: Fisher-Whitened Cross-Covariance for Low-Resource LLM Adaptation
Large Language Models
Efficient ML
Optimization
- FCCA optimizes a small core of parameters while fixing the bases of weight matrices, enhancing parameter-efficient fine-tuning.
- The method significantly outperforms existing low-resource adaptation techniques, achieving higher accuracy with fewer trainable parameters.
- Controlled experiments highlight the importance of task-relevant subspace selection and the benefits of whitening in improving adaptation performance.
- FCCA maintains competitive performance compared to methods like LoRA and DoRA while optimizing a fraction of the parameters.
Read more
Frozen Cores Need Task Signal: Fisher-Whitened Cross-Covariance for Low-Resource LLM Adaptation
Summary
This paper addresses the challenge of parameter-efficient fine-tuning (PEFT) in low-resource settings, focusing on the importance of the location of trainable parameters in addition to their quantity. The authors introduce a novel approach called Fisher-Whitened Cross-Covariance Adaptation (FCCA), which optimizes a small core of parameters while keeping the bases of the weight matrices fixed. This method allows for a clearer evaluation of the quality of the subspace used for adaptation. The FCCA method involves estimating the signed input-error cross-covariance, whitening it using diagonal Fisher moments, and applying thin QR decomposition to obtain stable core coordinates. The authors conduct extensive experiments across 11 tasks and multiple model settings, demonstrating that FCCA outperforms existing methods in terms of accuracy while significantly reducing the number of trainable parameters. The findings reveal that a carefully selected fixed span can recover much of the performance benefits typically associated with more complex low-rank adaptations, thus offering a more efficient alternative for model adaptation in constrained environments.
Methodology
The methodology involves a three-step process: (1) Calibration to estimate the signed cross-covariance and Fisher moments, (2) Selection of a subspace using a rank-constrained approach in Fisher-scaled coordinates, and (3) Training of the core parameters using thin QR decomposition to ensure stable optimization. This approach isolates the effects of subspace selection from other factors in parameter adaptation.
Results
FCCA achieved an 83.0 macro-average accuracy on the Qwen2.5-3B model, outperforming the next-best matched-budget constructor by 2.3 points and exceeding the performance of the unwhitened RawGrad control across all tasks. It ranked first on all Qwen scales and was nearly tied for first on Llama-3.2-1B. Controlled ablation studies indicated that whitening and QR decomposition were critical for stable core optimization, with performance gains ranging from 2.7 to 17.2 points.
Implications
The findings suggest that efficient model adaptation can be achieved with fewer trainable parameters by carefully selecting fixed spans in the parameter space. This has significant implications for deploying large language models in resource-constrained environments, enabling better performance without the need for extensive computational resources.
Explore More, Drift Less: Outcome-Only Reinforcement Learning Can Suffice for Long-Horizon Interactive Agents
Reinforcement Learning
Large Language Models
- The paper identifies signal starvation and policy drift as key issues limiting the effectiveness of outcome-only RL.
- CANOPY is introduced as a novel protocol that enhances exploration and maintains policy integrity.
- The proposed method achieved top performance on the AppWorld benchmark without additional supervision or complex scaffolding.
- The findings suggest that smaller models can effectively learn long-horizon tasks through outcome-only RL.
Read more
Explore More, Drift Less: Outcome-Only Reinforcement Learning Can Suffice for Long-Horizon Interactive Agents
Summary
This paper addresses the limitations of outcome-only reinforcement learning (RL) in training long-horizon interactive agents, particularly in the context of small open models. The authors argue that the perceived ceiling on performance is due to two main issues: signal starvation and policy drift. Signal starvation occurs when sparse rewards lead to ineffective learning signals, while policy drift results from repeated updates on a limited task pool, causing the policy to degrade. To overcome these challenges, the authors propose CANOPY (Coverage-ANchored On-PolicY RL), a protocol that enhances exploration and maintains policy integrity by focusing on same-task exploration and on-policy updates. The effectiveness of CANOPY is demonstrated through experiments on AppWorld, a benchmark for interactive coding tasks, where a Qwen3-14B policy trained using this method achieved top performance without the need for task-specific supervision or auxiliary signals. The results indicate that agentic RL can effectively internalize long-horizon capabilities in smaller models, challenging the notion that outcome-only RL is inherently limited.
Methodology
The authors developed CANOPY, which focuses on large same-task exploration to restore learning signals and employs KL-anchored, on-policy updates to prevent policy drift. This approach allows for an increased interaction budget during testing, enabling the agent to perform better in long-horizon tasks.
Results
The Qwen3-14B policy trained with CANOPY achieved a score of 86.9 on the Test-Normal TGC and 67.6 on the Test-Challenge of the AppWorld benchmark, outperforming other methods and demonstrating the effectiveness of the proposed approach. Additionally, the same principles improved the performance of Qwen3.5-9B on SWE-bench by 16.6 points.
Implications
The results indicate that outcome-only RL can be a viable approach for training interactive agents, particularly in applications requiring long-horizon decision-making. This could lead to more efficient training methods for smaller models in various domains, including software engineering and automation.
Scaled Idempotence in Transformer Attention: Paired OV Geometry and Shared-Value Algebras
NLP
Large Language Models
Theory
- Identified a recurrent scaled-idempotent behavior in Transformer attention heads across various model sizes.
- Demonstrated that strong closure alignment disappears under O/V mismatches, indicating a specific learned relationship.
- Separated geometric capacity from trained attainment, showing high closure is feasible even in layers without strong heads.
- Proved that shared values transform headwise closure into a fixed-gain operator algebra, confirmed through extensive model testing.
Read more
Scaled Idempotence in Transformer Attention: Paired OV Geometry and Shared-Value Algebras
Summary
This paper investigates a recurrent algebraic regularity in Transformer attention mechanisms, specifically focusing on a phenomenon termed 'scaled idempotence.' The authors identify that a sparse subset of effective output-value (OV) operators in Transformer models nearly closes under composition, leading to the relation T² ≈ αT. This behavior is observed across various pretrained models with parameters ranging from 2.8 billion to 235 billion. The study reveals that approximately 3.98% to 8.00% of attention heads exhibit strong squared closure alignment, while no matched within-layer OV mismatches reach similar thresholds. The authors introduce a factorization in principal coordinates that separates the transport geometry from the read-write return geometry, allowing for a deeper understanding of the mechanisms at play. They conduct extensive experiments across multiple models and layers, demonstrating that scrambling the trained orientation of the OV operators significantly reduces closure, while preserving other properties. The findings suggest that scaled idempotence is a signature of a paired O/V relation rather than a property of individual factors, and that shared values can extend this relationship into a local operator algebra.
Methodology
The authors employed a combination of algebraic analysis and empirical testing across multiple pretrained Transformer models. They analyzed the effective OV operators, decomposed them into principal-angle geometry, and conducted interventions to assess the impact of scrambling trained orientations. The study involved examining closure properties and conducting experiments on 7,304 attention heads across nine models.
Results
The study found that a small percentage of attention heads exhibited strong squared closure alignment, with median closure values significantly dropping when trained orientations were scrambled. The authors established a factorization that separates the transport core from the read-write geometry, confirming that the trained orientation plays a crucial role in achieving high closure. Additionally, they demonstrated that shared values enable a fixed-gain operator algebra among attention heads.
Implications
The findings have significant implications for understanding the inner workings of Transformer models, particularly in how attention mechanisms can be optimized and interpreted. The insights into scaled idempotence and the relationship between output and value factors could inform future model architectures and training strategies, potentially leading to more efficient and interpretable models.
HarnessEvolve: Learning from Reference Trajectories for Reliable Agent Self-Evolution
NLP
Large Language Models
Optimization
- HarnessEvolve decouples execution and optimization into independent modules for better reliability.
- The framework generates reference trajectories to improve error signal extraction and systematic failure identification.
- Quality and performance gates are implemented to prevent shortcut learning and catastrophic forgetting.
- Extensive experiments show consistent performance improvements over existing self-evolving frameworks.
Read more
HarnessEvolve: Learning from Reference Trajectories for Reliable Agent Self-Evolution
Summary
The paper presents HarnessEvolve, a novel self-evolving framework designed to enhance the autonomy of agents by optimizing their harness—comprising prompts, skills, tools, and execution logic—based on environmental feedback. The authors identify three primary challenges in current self-evolving agents: credit assignment failure, shortcut learning, and catastrophic forgetting. HarnessEvolve addresses these challenges by decoupling the execution agent from the evolutionary pipeline, allowing for independent modules for execution, evaluation, optimization, and gating. This modular approach enables the generation of reference trajectories to extract error signals from failed executions, which are then clustered to identify systematic failure patterns. To mitigate shortcut learning and catastrophic forgetting, candidate harness updates must pass through quality and performance gates. Extensive experiments across various benchmarks demonstrate that HarnessEvolve consistently outperforms state-of-the-art baselines, confirming its reliability and effectiveness in diverse task domains.
Methodology
HarnessEvolve employs a modular architecture where execution, evaluation, optimization, and gating are handled by separate agents. The execution agent performs tasks, while the evaluation agent assesses accuracy and reliability. The optimization agent analyzes failed executions against reference trajectories to identify root causes of errors, and the gate agent filters candidate updates for quality and performance.
Results
HarnessEvolve was tested on five benchmarks, including open-domain tasks and specialized enterprise scenarios. It consistently outperformed existing self-evolving baselines in terms of accuracy, demonstrating its effectiveness across various settings.
Implications
The findings suggest that HarnessEvolve can significantly improve the reliability and autonomy of self-evolving agents in real-world applications, potentially leading to advancements in fields such as automated software engineering, enterprise data analysis, and other complex task domains.
Elite-Weighted Supervised Fine-tuning for Goal-Directed Molecular Optimization
Generative Models
Optimization
- EW-SFT allows for effective goal-directed updates across diverse molecular generator architectures and tasks.
- The method emphasizes elite selection of high-scoring molecules as the primary channel for incorporating reward information.
- EW-SFT achieves superior performance compared to traditional policy-gradient methods in molecular optimization tasks.
- The approach is applicable to both de novo and fragment-constrained molecular generation.
Read more
Elite-Weighted Supervised Fine-tuning for Goal-Directed Molecular Optimization
Summary
This paper introduces Elite-Weighted Supervised Fine-tuning (EW-SFT), a novel approach for goal-directed molecular optimization that enhances the performance of molecular generators. Traditional methods often rely on policy-gradient reinforcement learning (RL), which can be cumbersome due to architecture-specific requirements and the need for trajectory log-probabilities. In contrast, EW-SFT leverages a supervised fine-tuning framework that incorporates reward information through elite selection of high-scoring molecules, allowing for updates based solely on the model's native pretraining loss. This method is versatile, applicable across various molecular generator architectures and design tasks, including de novo generation and motif-extension. The authors demonstrate that EW-SFT consistently outperforms existing native optimizers in multiple scenarios, including ligand-based molecular design, achieving significant improvements in similarity metrics while maintaining efficiency without the complexities of RL. The findings suggest that EW-SFT is a robust and unified optimization strategy for molecular design, capable of adapting to different generation procedures and constraints.
Methodology
The methodology involves using a rolling elite buffer to retain high-scoring molecules, which are then selected for fine-tuning the model using its native pretraining loss. This process eliminates the need for trajectory-level log-probabilities and allows for updates based solely on scored molecules, making it adaptable across different molecular generation architectures.
Results
EW-SFT demonstrated consistent improvements in Top-1k similarity metrics over native optimizers under a fixed budget of 3D shape alignment oracle calls. It also showed enhanced performance in 2D similarity optimization tasks and comparable results to tuned policy-gradient methods on the Practical Molecular Optimization benchmark.
Implications
The implications of this research extend to the field of drug discovery and molecular design, where efficient and effective optimization of molecular candidates is crucial. EW-SFT could streamline the process of generating novel compounds with desired properties, potentially accelerating the development of new therapeutics.
HarmoCore: Functional Latent Diffusion for Sparse Reconstruction of Oscillatory Wave Fields
Generative Models
Optimization
Theory
- Introduction of HarmoCore, a functional latent diffusion framework for reconstructing oscillatory wave fields.
- Utilization of Functional Tucker models for compact representation of complex-valued wave fields.
- Frequency-conditioned diffusion model trained in core space, enhancing efficiency and accuracy.
- Demonstrated substantial improvements in reconstruction performance under extreme sparsity.
Read more
HarmoCore: Functional Latent Diffusion for Sparse Reconstruction of Oscillatory Wave Fields
Summary
The paper presents HarmoCore, a novel framework designed for the sparse reconstruction of oscillatory wave fields from limited sensor measurements. Traditional methods for reconstructing physical fields often struggle with the complexities of oscillatory, complex-valued data, especially when faced with extreme sparsity in observations. HarmoCore addresses these challenges by employing a generative prior in a compact, continuous, and structured latent space. The framework utilizes Functional Tucker cores to represent joint real and imaginary channels of wave fields, allowing for efficient frequency-conditioned diffusion modeling. By performing Diffusion Posterior Sampling directly in core space, the method avoids the inefficiencies associated with dense pixel-space corrections. The authors demonstrate the effectiveness of HarmoCore through experiments on various wave field scenarios, achieving significant improvements in reconstruction accuracy even with as little as 1%–2% sensing. This work highlights the potential of using advanced latent representations and diffusion techniques for complex wave field reconstruction tasks.
Methodology
HarmoCore employs a functional latent diffusion approach, representing wave fields using Functional Tucker cores that capture joint real and imaginary components. The framework trains a frequency-conditioned diffusion model in core space and utilizes a multilinear decoder to derive observation-likelihood gradients and residual guidance efficiently. This allows for direct sampling and reconstruction from sparse sensor data without the need for dense pixel-space corrections.
Results
The experiments conducted on 2D Helmholtz, 2D synthetic wave fields, and 3D Helmholtz scenarios show that HarmoCore achieves significant reconstruction improvements, even with extreme sparsity in sensor measurements (1%–2%). The method demonstrates its capability to maintain physical consistency and coherence in the reconstructed wave fields.
Implications
The findings suggest that HarmoCore could be applied in various fields requiring accurate wave field reconstruction from limited data, such as electromagnetic simulations, ocean acoustics, and geophysical imaging. The approach may also inspire further research into efficient reconstruction techniques for other complex-valued fields.
A hybrid quantum-classical neural network for learning to route
Optimization
- Hybrid quantum-classical neural networks can effectively reduce model parameters while maintaining performance.
- The encoder feed-forward replacement strategy is identified as a promising approach for hybrid-module compression.
- Classical routing algorithms remain competitive and often superior to the hybrid model on larger instances.
- The study highlights the limitations of current quantum devices in fully replacing classical neural architectures.
Read more
A hybrid quantum-classical neural network for learning to route
Summary
This paper investigates the use of hybrid quantum-classical neural networks to develop routing heuristics, specifically focusing on the capacitated vehicle routing problem (CVRP). The authors explore whether small quantum neural networks (QNNs) can effectively replace parameter-heavy components of a classical attention-based routing model without sacrificing solution quality. The study finds that replacing the encoder feed-forward component with a quantum module can reduce the number of model parameters by 56.6% while maintaining competitive performance against classical neural networks for small to medium-sized instances. However, the performance gap widens for larger instances. The paper also compares the hybrid model's performance with classical routing algorithms, which often outperform the hybrid approach on fixed Euclidean test sets. The findings suggest that while there is no clear quantum advantage or solver dominance, the encoder feed-forward replacement strategy presents a viable method for compressing neural combinatorial optimization models.
Methodology
The authors propose a hybrid quantum-classical neural network architecture that incorporates a quantum neural network module to replace the encoder feed-forward component of a classical routing model. The model is evaluated on the capacitated vehicle routing problem (CVRP) using various instance sizes, and its performance is compared against classical neural networks and traditional routing algorithms.
Results
The hybrid quantum-classical model achieves a 56.6% reduction in parameters while maintaining competitive performance for small and medium instance sizes. However, the performance gap increases for larger instances, indicating that classical approaches still hold an advantage in those scenarios.
Implications
The findings suggest potential pathways for integrating quantum computing into neural network architectures, particularly in combinatorial optimization tasks. The research may inform future developments in quantum machine learning and optimization strategies.
Safin-1: Safety from Within through Memory-Native State Evolution
Large Language Models
NLP
Theory
- Safin-1 integrates safety as an intrinsic property of the model through memory-native state evolution.
- The architecture employs Memory-Anchor Routing to maintain structured memory states and enable selective retrieval.
- Test-time adaptation allows for persistent safety capabilities without altering the backbone model.
- Significant safety improvements were observed in evaluations compared to existing models.
Read more
Safin-1: Safety from Within through Memory-Native State Evolution
Summary
The paper introduces Safin-1, a family of foundation models designed to enhance safety in long-horizon complex tasks through a novel architecture called Memory-Anchor Routing across Context History (MARCH). Unlike traditional models that rely on external safeguards or post-hoc adjustments for safety, Safin-1 integrates safety as an intrinsic property of the model itself. This is achieved by allowing the model to maintain structured memory states and adapt its capabilities over time through a persistent Safety State. The architecture supports selective retrieval of historical information, enabling the model to evolve its behavior based on context while ensuring safety. The authors demonstrate that Safin-1 can effectively adapt its safety capabilities during test time without modifying the core model, leading to significant improvements in safety performance. The findings suggest a new paradigm where model memory serves as an active substrate for evolving behavior, paving the way for future research in safety and long-horizon intelligence.
Methodology
The authors developed the Safin-1 architecture based on MARCH, which facilitates the preservation of historical states as addressable internal memory. The model uses a routing interface to retrieve relevant states dynamically and incorporates a detachable Safety State for persistent capability specialization. The architecture was validated through experiments focusing on language modeling, long-context understanding, and safety evaluations.
Results
Evaluations showed that Safin-1 achieved substantial improvements in safety metrics, with a 47% reduction in average safety risk and an 11.4% increase in reasoning capabilities compared to baseline models. The architecture demonstrated effective long-context understanding and retrieval capabilities, validating its efficiency and scalability.
Implications
The findings suggest that integrating safety into the core architecture of foundation models can lead to more reliable AI systems capable of handling complex tasks over extended interactions. This approach could influence future designs of AI models, emphasizing the importance of safety as a fundamental characteristic rather than an add-on feature.
Higher Structures in Deep Learning
Theory
Optimization
Generative Models
- Higher-arity tensor operations are essential for understanding deep learning models.
- The paper introduces neural hypernetworks as a generalization of multilayer perceptrons.
- Connections between higher structures and evolutionary algorithms are explored.
- Current deep learning frameworks often neglect the potential of higher-order tensor interactions.
Read more
Higher Structures in Deep Learning
Summary
This paper explores the significance of higher-arity tensor operations in deep learning, addressing the limitations of current models that primarily rely on binary tensor operations. The authors argue that understanding the internal mechanics of deep learning models is crucial for improving their interpretability and performance. They introduce a novel empirical investigation into higher-arity phenomena within trained neural networks and propose a hypergraphical generalization of the multilayer perceptron. This approach allows for a more comprehensive analysis of tensor interactions, which are often overlooked in traditional deep learning frameworks. The paper also discusses the potential connections between these higher structures and evolutionary algorithms, suggesting that such relationships could lead to new insights and advancements in model architecture. The authors conclude by outlining promising directions for future research in this area, emphasizing the need for a more principled mathematical foundation in deep learning.
Methodology
The authors conducted an empirical investigation of higher-arity phenomena in neural networks, leveraging symbolic and compositional representations of tensor operations. They introduced the concept of neural hypernetworks and explored connections to evolutionary algorithms to analyze the implications of higher-order structures in deep learning.
Results
The investigation revealed significant higher-arity phenomena in logit trajectories of neural networks. The introduction of neural hypernetworks demonstrated a new way to generalize multilayer perceptrons, highlighting the untapped potential of higher-order tensor operations in enhancing model architectures and interpretability.
Implications
The findings suggest that incorporating higher-arity tensor operations could lead to more powerful and interpretable deep learning models. This could have broad applications across various domains, including natural language processing, computer vision, and reinforcement learning, by improving model performance and understanding.
WHALE: A Simple Recipe for Joint Harness-Weight Optimization
Optimization
Large Language Models
Reinforcement Learning
- WHALE alternates between updating model weights and searching for better harness code, addressing performance bottlenecks.
- The framework uses online rejection-sampling fine-tuning and Meta-Harness for its two phases.
- WHALE outperforms traditional weight-only and harness-only optimization methods by 4.15–24.38 percentage points in accuracy.
- Adaptive phase switching based on training signals enhances performance compared to fixed schedules.
Read more
WHALE: A Simple Recipe for Joint Harness-Weight Optimization
Summary
The paper introduces WHALE (Weight-Harness Alternating LEarning), a novel framework for optimizing both model weights and harness code in agent-based systems. Traditional methods often optimize these components in isolation, leading to performance bottlenecks. WHALE alternates between two phases: updating the model weights under the current harness and searching for a better harness with the updated model. The authors implement this approach using online rejection-sampling fine-tuning for weight updates and Meta-Harness for harness search. They address the challenge of timing for switching between phases by proposing fixed durations or an adaptive patience rule based on training signals. The framework is evaluated across three domains: search question answering, mathematical reasoning, and chess puzzles. Results show that WHALE significantly outperforms single-component optimization methods, demonstrating the importance of joint optimization in improving agent performance.
Methodology
WHALE employs a two-phase alternating approach: first, it updates model weights using online rejection-sampling fine-tuning, and second, it searches for an improved harness using Meta-Harness. The timing for switching between these phases can be fixed or adaptive based on the training signals.
Results
WHALE demonstrated improvements in accuracy ranging from 4.15 to 24.38 percentage points over weight-only and harness-only methods across three domains. In SearchQA, harness search matched peak weight-only accuracy with fewer rollouts, while in Math, it required a weight update to be effective. Adaptive WHALE outperformed fixed schedules in both evaluated domains.
Implications
The findings suggest that joint optimization of model weights and harness code can lead to significant performance gains in agent-based systems. This approach could be applied to various domains where agents interact with complex environments, enhancing their adaptability and effectiveness.
Predicting Subsurface Abnormalities Growth using Physics-Informed Neural Networks
Theory
Time Series
Interpretability
- Integration of Physics-Informed Neural Networks (PINNs) into GPR data prediction.
- Development of a specialized model combining CNN, SFCA, and ConvLSTM for enhanced accuracy.
- Improved reliability and interpretability of nondestructive evaluation (NDE) results.
- Potential transformation of infrastructure condition assessments through physics-based modeling.
Read more
Predicting Subsurface Abnormalities Growth using Physics-Informed Neural Networks
Summary
This research introduces a novel application of Physics-Informed Neural Networks (PINNs) to enhance the prediction of subsurface abnormalities using Ground-Penetrating Radar (GPR) data. Drawing parallels with medical imaging techniques for tumor tracking, the study develops a specialized PINN model that integrates deep learning with the physical principles governing electromagnetic wave propagation. The architecture comprises a Convolutional Neural Network (CNN), a spatial feature channel attention (SFCA) mechanism, and ConvLSTM with temporal feature frame attention (TFFA) modules. This model effectively captures spatial and temporal patterns in GPR data, improving the accuracy of predictions compared to traditional methods that often rely on isolated data points. By embedding physics directly into the neural network, the model not only aligns with fundamental physical laws but also enhances the interpretability and reliability of nondestructive evaluation (NDE) results for infrastructure assessments. The findings indicate that this approach can significantly improve the assessment of bridge deck conditions, providing deeper insights into deterioration mechanisms and facilitating better maintenance strategies.
Methodology
The study employs a Physics-Informed Neural Network (PINN) framework that combines deep learning techniques with the physics of electromagnetic wave propagation. The architecture includes a CNN for feature extraction, a spatial feature channel attention mechanism for focusing on relevant spatial features, and ConvLSTM modules to capture temporal dependencies in the data. The model is trained on GPR data to predict subsurface abnormalities, leveraging both spatial and temporal contextual information.
Results
The proposed PINN model demonstrated significant improvements in forecasting GPR data accuracy compared to traditional methods. The integration of physics into the neural network architecture allowed for predictions that are not only more precise but also aligned with fundamental physical principles. This advancement is crucial for effective assessments of bridge deck conditions and other civil infrastructure evaluations.
Implications
The research suggests that incorporating Physics-Informed Neural Networks into nondestructive evaluation can enhance the precision of infrastructure deterioration predictions. This approach could lead to more effective maintenance and repair strategies, ultimately improving the longevity and safety of civil infrastructure. Additionally, it offers a framework for further exploration of physics-based models in other domains of engineering and diagnostics.
DISTAL: Distillation and Self-Supervised Pretraining for Structure-Agnostic Materials Property Prediction
Graph Learning
- DISTAL combines self-supervised pretraining and knowledge distillation for materials property prediction.
- The framework operates without requiring structural information during inference, making it suitable for early-stage screening.
- It significantly outperforms existing benchmarks across multiple tasks, showcasing the effectiveness of integrating compositional and structural knowledge.
- The approach addresses the challenges of low-data settings in materials informatics.
Read more
DISTAL: Distillation and Self-Supervised Pretraining for Structure-Agnostic Materials Property Prediction
Summary
The paper presents DISTAL, a novel framework aimed at improving materials property prediction in scenarios with limited labeled data. Traditional models often rely on crystal structures, which can be unavailable during early-stage screening. DISTAL addresses this limitation by integrating self-supervised compositional pretraining with structure-aware knowledge distillation. The framework first learns transferable compositional representations from a large virtual composition space using 145 composition-derived descriptors. It then distills structural knowledge from a pretrained ALIGNN teacher model into a composition-conditioned student model. This allows the model to leverage structural priors during training without needing structural inputs during inference. The authors demonstrate that by combining explicit compositional descriptors, pretrained latent features, and distilled structural features, DISTAL captures complementary signals that enhance predictive performance. The results show that the best-performing multimodal configuration improves upon reference benchmarks in 37 out of 39 tasks, indicating that the integration of compositional pretraining and structural distillation provides a robust approach for composition-only prediction in materials informatics.
Methodology
DISTAL employs a dual-prior framework that includes self-supervised learning to derive compositional representations from a large dataset of composition-derived descriptors and knowledge distillation from a pretrained ALIGNN model to incorporate structural insights. This combination allows the model to learn effective representations without relying on structural data during inference.
Results
The DISTAL framework was evaluated across 39 benchmark tasks, achieving the best performance in 37 of them compared to reference benchmarks. The integration of compositional descriptors, pretrained features, and distilled structural knowledge led to significant improvements in predictive accuracy.
Implications
The findings suggest that DISTAL can facilitate more efficient materials discovery by enabling robust predictions in low-data environments, potentially accelerating the identification of functional materials in various applications, including electronics, energy storage, and catalysis.
A Study of Hidden-State Optimization Order in Predictive Coding Networks
Optimization
Theory
Efficient ML
- Introduces a boundary-first hidden-state optimization schedule for local-error training systems.
- Demonstrates significant accuracy improvements (9.77% on standard parametrization) over standard predictive coding.
- Enables all layers to receive non-trivial updates, enhancing feature learning.
- Provides diagnostic analyses showing reduced CKA and increased gradient diversity.
Read more
A Study of Hidden-State Optimization Order in Predictive Coding Networks
Summary
This paper investigates the optimization order of hidden states in Predictive Coding Networks (PCNs) to enhance feature learning, which is often limited by local learning methods. The authors propose a boundary-first inference schedule that organizes the optimization process into chunks, first aligning hidden states at chunk boundaries before refining the representations within each chunk. This approach aims to address the weaknesses of standard PCNs, which have been shown to produce weak updates in early layers due to imbalanced error propagation. The proposed method, referred to as PC-PC, is evaluated on the CIFAR-10 dataset, demonstrating significant improvements in accuracy over standard predictive coding methods. The authors also analyze the representation learning dynamics, revealing that their approach leads to more informative updates across all layers and promotes stronger feature learning. The findings suggest that the boundary-first optimization order is a promising design principle for local-learning systems and could be extended to other frameworks in future research.
Methodology
The authors propose a boundary-first inference schedule that partitions the model into chunks, aligning hidden states at chunk boundaries through inter-chunk inference before refining the states within each chunk via intra-chunk inference. This method is instantiated in predictive coding networks, allowing for the explicit exposure of hidden activities and prediction errors during inference.
Results
The boundary-first predictive coding instantiation (PC-PC) improves accuracy on CIFAR-10 by 9.77% under standard parametrization and by 5.51% under a µ-parametrization. Diagnostic analyses reveal that all layers receive non-trivial updates, with a reduction in initial-to-final CKA and increased diversity in layerwise gradients compared to standard predictive coding and global backpropagation.
Implications
The findings suggest that optimizing the order of hidden-state updates can significantly enhance feature learning in predictive coding networks and potentially in other local-learning systems. This could lead to more effective training methodologies that are biologically plausible and modular, with applications in various machine learning domains.
The Multiple Timescales of Gradient Descent on the Edge of Stability: A Perturbative Derivation of the Central Flow
Optimization
Theory
- Introduces a perturbative regime to derive the central flow of gradient descent.
- Identifies three distinct timescales in the dynamics of gradient descent.
- Uses the method of multiple scales to formalize the approximation of gradient descent.
- Provides insights into self-stabilization mechanisms in optimization.
Read more
The Multiple Timescales of Gradient Descent on the Edge of Stability: A Perturbative Derivation of the Central Flow
Summary
This paper addresses the theoretical understanding of gradient descent in deep learning, particularly at the edge of stability where the sharpness of the loss function fluctuates. The author critiques the heuristic derivation of the central flow proposed by Cohen et al. (2025) and introduces a perturbative regime to formalize this derivation. By assuming a loss function that decomposes into two parts, the author shows that as the perturbation parameter approaches zero, the dynamics of gradient descent converge to the gradient flow of the perturbation constrained to minimizers of the main loss function. The analysis reveals three distinct timescales: a fast oscillation along the sharpest direction, an intermediate self-stabilization mechanism, and a slow evolution along the minimizers of the main loss function. The method of multiple scales is employed to derive the dynamics, leading to the central flow as the leading-order term. This work not only provides a more rigorous foundation for the central flow approximation but also deepens the understanding of self-stabilization mechanisms in gradient descent, especially in scenarios with multiple eigenvalues at the edge of stability.
Methodology
The author employs a perturbative approach, treating gradient descent as a singularly perturbed dynamical system. The method of multiple scales is utilized to derive an approximation that separates the dynamics into three timescales, allowing for a clearer understanding of the central flow and self-stabilization mechanisms.
Results
The analysis confirms that the central flow emerges as the leading-order term in the perturbative expansion of gradient descent dynamics. It also elucidates the role of self-stabilization in maintaining stability during optimization, particularly when multiple eigenvalues are present at the edge of stability.
Implications
This work enhances the theoretical framework for understanding gradient descent in deep learning, potentially leading to improved optimization strategies and insights into the behavior of neural networks during training. It also opens avenues for further research into the dynamics of optimization in high-dimensional spaces.
iPINN for Broadband CARS Phase Retrieval: A Framework for Function Approximation and Inverse Modeling Problems in Nonlinear Spectroscopy
Theory
- Introduction of iPINN for robust phase retrieval in BCARS.
- Utilization of a transformer encoder for spectral feature assignment.
- Achievement of the lowest MAE on a public benchmark compared to existing methods.
- Demonstration of depth-invariant accuracy across various solvents.
Read more
iPINN for Broadband CARS Phase Retrieval: A Framework for Function Approximation and Inverse Modeling Problems in Nonlinear Spectroscopy
Summary
The paper presents an innovative approach to phase retrieval in broadband coherent anti-Stokes Raman spectroscopy (BCARS) using an inverse physics-informed neural network (iPINN). This method addresses the challenges of recovering resonant susceptibility from raw BCARS spectra, which is complicated by the presence of a non-resonant background (NRB) that varies across different acquisitions. The iPINN framework predicts Lorentzian peak parameters directly from the spectra and reconstructs the resonant susceptibility using a differentiable analytical forward model. A transformer encoder is employed to assign spectral features to 24 learnable peak slots, while a multi-view consistency loss ensures robustness against variations in NRB patterns, strengths, and noise. The proposed method outperforms traditional spectral regression techniques, achieving the lowest mean absolute error (MAE) on a public benchmark and demonstrating depth-invariant accuracy across multiple solvents and focal positions. The results indicate that the iPINN framework effectively supports robust phase retrieval in BCARS, making it a significant advancement in the field of nonlinear spectroscopy.
Methodology
The iPINN framework employs a transformer encoder to map raw BCARS spectra to a compact set of Lorentzian peak parameters. It incorporates a differentiable physics layer to reconstruct the resonant signal from these parameters, ensuring that the model is robust against variations in NRB and noise through a multi-view consistency loss.
Results
The iPINN achieved a mean absolute error (MAE) of 0.016 on a public benchmark, significantly lower than the next best result of 0.046. Additionally, it demonstrated depth-invariant accuracy across five out of seven solvents in zero-shot tests, indicating its robustness under varying measurement conditions.
Implications
The iPINN framework has the potential to enhance the accuracy and reliability of phase retrieval in nonlinear spectroscopy, which could lead to improved chemical analysis and diagnostics in various scientific fields, including material science and biophysics.
Breaking the Structural Identity: Personalized Federated LoRA Fine-tuning under Rank Heterogeneity
NLP
Large Language Models
Federated Learning
- FedRoRA enables fine-grained personalization in federated learning under rank heterogeneity.
- The framework decouples adaptation into shared global directions and personalized rank-wise magnitudes.
- A personalized aggregation mechanism allows clients with the same rank to receive distinct updates based on local task alignment.
- Extensive experiments show that FedRoRA outperforms existing rank-heterogeneous federated learning methods.
Read more
Breaking the Structural Identity: Personalized Federated LoRA Fine-tuning under Rank Heterogeneity
Summary
This paper addresses the challenges of adapting Large Language Models (LLMs) to privacy-sensitive, distributed datasets through a novel framework called FedRoRA (Federated Rank-wise Personalized LoRA). The authors identify that existing federated learning methods, particularly those utilizing Low-Rank Adaptation (LoRA), struggle with rank and data heterogeneity, leading to suboptimal performance in non-IID scenarios. FedRoRA introduces a decoupled approach to fine-tuning, separating adaptation directions and rank-wise magnitudes, allowing for personalized updates even among clients with the same rank. The framework employs singular value decomposition (SVD) to extract a global subspace and redistributes client-specific initializations through a personalized projection mechanism. Extensive experiments on NLU and NLG benchmarks demonstrate that FedRoRA consistently outperforms existing state-of-the-art methods, highlighting its effectiveness in addressing the dual challenges of resource and data heterogeneity in federated settings.
Methodology
The authors propose FedRoRA, which replaces the standard LoRA update mechanism with a decoupled triplet structure. It uses singular value decomposition (SVD) to extract a shared global subspace from local updates and redistributes personalized triplets for client-specific initialization. This approach allows for distinct adaptation directions and magnitudes tailored to each client's local data characteristics.
Results
The experimental results on GLUE and FLAN benchmarks indicate that FedRoRA consistently surpasses state-of-the-art rank-heterogeneous federated learning methods, particularly in non-IID settings, demonstrating its robustness and effectiveness in personalized federated fine-tuning.
Implications
The findings suggest that FedRoRA can significantly enhance the performance of federated learning systems in real-world applications where data privacy and resource constraints are critical. This framework could be applied in various domains requiring collaborative learning without compromising data security.
Modelpedia: A Catalog of Model Findings for the Meta-Science of AI
Theory
Interpretability
Large Language Models
- Modelpedia serves as a centralized knowledge catalog for findings related to AI models.
- The framework utilizes an automated, LLM-assisted pipeline for data extraction and aggregation.
- Over a thousand findings were extracted from ICLR 2024 and 2025 papers, demonstrating the framework's capability.
- The catalog allows for meta-analysis of community research focus, enhancing understanding of model behaviors.
Read more
Modelpedia: A Catalog of Model Findings for the Meta-Science of AI
Summary
The paper introduces Modelpedia, an automated framework designed to catalog findings about AI models, addressing the challenge of scattered knowledge in the rapidly evolving field of AI. As new foundation models emerge, a plethora of research documents their behaviors, failures, and biases, but these findings remain largely unorganized and difficult to retrieve. Modelpedia aims to consolidate this information into a searchable public catalog, linking findings to specific models, datasets, methods, and concepts. The authors implemented a prototype that successfully extracted over a thousand findings from papers accepted to ICLR 2024 and 2025, and conducted a meta-analysis to understand community focus areas. The framework not only serves as a repository but also encourages community contributions, fostering a shared foundation for the meta-science of AI.
Methodology
The authors developed an automated, LLM-assisted pipeline that extracts findings from published papers, organizes them into a structured database, and links each finding to relevant models, datasets, methods, and concepts. The process involves ranking papers for relevance, extracting pertinent data, and updating the Modelpedia catalog.
Results
The prototype implementation of Modelpedia successfully extracted and cataloged over a thousand findings from recent ICLR papers. A meta-analysis of these findings revealed insights into community research trends and highlighted areas of concentrated attention regarding model behaviors and performance.
Implications
Modelpedia has the potential to significantly enhance the organization and accessibility of knowledge in AI research, facilitating better understanding and comparison of model behaviors. It encourages systematic investigation of models, which could lead to improved model development and deployment practices in the AI community.
Superposed Latent Autoencoder
Computer Vision
Efficient ML
Generative Models
- Introduction of SLAE, which allows multiple wide latents to share memory instead of shrinking each independently.
- Formulation of superposed latent storage as a capacity–interference tradeoff.
- Demonstrated significant improvements in reconstruction error and downstream classification accuracy across multiple datasets.
Read more
Superposed Latent Autoencoder
Summary
The paper introduces the Superposed Latent Autoencoder (SLAE), a novel architecture designed to enhance the efficiency of latent representation storage in autoencoders. Traditional autoencoders typically reduce the dimensionality of latent representations to fit within tight memory budgets, which often leads to a loss of representational capacity. SLAE, in contrast, allows for multiple wider latent representations to be stored together through a process of learned superposition. This method transforms latents into storage-friendly codes, binds them with randomized keys, and superposes these codes into a single memory tensor. The architecture learns to recover individual latents from this shared memory, effectively replacing the dimensionality reduction bottleneck with a structured interference that can be managed. The authors demonstrate that SLAE significantly improves the reconstruction-memory tradeoff across various datasets, achieving up to a 56% reduction in reconstruction error compared to conventional autoencoders while maintaining the same storage budget. Additionally, the preserved information enhances downstream classification tasks, yielding improvements of up to 16.79 percentage points. The findings suggest that maintaining wider representations and utilizing superposition can be a more effective strategy for representation compression than merely shrinking latents.
Methodology
SLAE encodes inputs into wider latent representations, transforms these into storage-friendly codes, and superposes multiple codes into a shared memory tensor. It utilizes randomized keys for binding and employs a learned recovery network to demix the shared memory for individual representation retrieval.
Results
SLAE achieves up to a 56% reduction in reconstruction error compared to traditional autoencoders at matched storage budgets. It also improves downstream classification accuracy by up to 16.79 percentage points under the same memory constraints.
Implications
The findings suggest a new paradigm for representation compression in machine learning, indicating that wider representations can be more effective than smaller ones when memory is limited. This approach could have applications in various fields requiring efficient data representation, such as image processing and classification tasks.
From Truncation to Commitment: Persistent Context in Uniform Discrete Diffusion
Generative Models
NLP
Large Language Models
- CRS introduces persistent context in discrete diffusion models, enhancing sequence generation.
- Theoretical analysis supports the design choices of CRS, including a warmup phase and visibility of selected tokens.
- Empirical results show CRS achieves lower generative perplexity compared to traditional top-p sampling methods.
- The distinction between support restriction and persistent context is crucial for understanding generative performance.
Read more
From Truncation to Commitment: Persistent Context in Uniform Discrete Diffusion
Summary
This paper introduces committed reveal sampling (CRS), a novel sampling method for uniform-state discrete diffusion models that enhances the generation of sequences by maintaining selected tokens as persistent context for future predictions. Unlike traditional top-p sampling, which restricts predictions to a limited set of candidates at each step, CRS allows for the selected tokens to remain visible in subsequent sampling steps. The author provides a theoretical foundation for this approach, demonstrating that the optimal selection error does not increase as noise decreases, and that keeping selected tokens visible can help align later predictions towards the same sequence-level choice. Empirical results show that CRS outperforms fixed top-p baselines in terms of generative perplexity (GenPPL) across various function evaluation budgets, indicating a more favorable tradeoff between GenPPL and unigram entropy. The findings suggest that persistent context and support restriction are distinct mechanisms that can be leveraged to improve generative performance in discrete diffusion models.
Methodology
The methodology involves a training-free sampling algorithm that utilizes selected clean-token hypotheses as persistent context. After an initial warmup phase with a native revisable sampler, CRS stores argmax tokens from concentrated predictions and incorporates them into future model inputs, while unselected positions continue to follow the native sampler. Theoretical analyses and empirical experiments are conducted to evaluate the performance of CRS against traditional top-p sampling methods.
Results
CRS demonstrates lower generative perplexity (GenPPL) than fixed top-p baselines (p = 0.95 and p = 0.9) across various budgets of function evaluations (NFE). At 64 NFE, CRS also shows a favorable GenPPL-entropy tradeoff, indicating improved performance in generating coherent sequences. The results highlight the effectiveness of persistent context in enhancing the generative capabilities of discrete diffusion models.
Implications
The findings suggest that incorporating persistent context into sampling methods can significantly improve the performance of generative models in various applications, such as natural language processing and sequence generation tasks. This approach could lead to advancements in model design and training strategies for better coherence and accuracy in generated outputs.
Context Window Failures in Relational Foundation Models
Graph Learning
Time Series
Theory
- Introduction of Animus, a synthetic dataset to test relational model limits.
- Current relational models struggle with high-cardinality data, leading to poor performance.
- Pre-aggregation of data significantly improves model performance.
- Highlights the inadequacy of existing models for real-world relational data challenges.
Read more
Context Window Failures in Relational Foundation Models
Summary
This paper addresses the limitations of recent relational deep learning architectures, which are designed to handle multi-table relational data but impose strict constraints on the number of rows that can be processed for each entity. The authors introduce Animus, a synthetic financial dataset that simulates the challenge of predicting customer income based on a vast number of transactions. The study evaluates three recent models (Relational Transformer, Griffin, and Relational Graph Transformer) and finds that they perform poorly on raw data due to their budget limitations, achieving an R2 score of 0.18. However, when a simple pre-aggregation step is applied to the data, the performance improves significantly to an R2 of 0.65. This raises concerns about the readiness of current relational foundation models for high-cardinality real-world data, as they tend to discard potentially informative records beyond their processing limits. The paper emphasizes the need for more robust models that can effectively manage high-frequency entities in relational databases.
Methodology
The authors created the Animus dataset, which includes 100,000 customers with varying transaction frequencies. They evaluated four models (RT, Griffin, RelGT, and GraphSAGE) on both raw and pre-aggregated data to assess their performance in predicting customer income. The models were tested under a temporal framework, with specific training and validation cutoffs.
Results
The evaluation revealed that the three budget-limited models (RT, Griffin, RelGT) achieved an R2 score of 0.18 on raw data, while a single pre-aggregation step improved their performance to 0.65. In contrast, GraphSAGE, which operates under a larger sampling budget, performed better on raw data and showed only a marginal improvement with pre-aggregation.
Implications
The findings suggest that current relational foundation models may not be suitable for high-cardinality data applications, such as financial transactions. This highlights the need for the development of more flexible models that can accommodate the complexities of real-world relational datasets, potentially impacting areas such as fraud detection and credit scoring.
MUGEN: Generating Unlearnable Graph Examples for Multiple Learning Tasks
Graph Learning
- MUGEN is the first framework to provide unified protection for multiple graph learning tasks with a single dataset.
- The Task-Aligned Separability Objective (TASO) enhances class separation in perturbation optimization, improving unlearnability.
- Type-Adaptive Perturbation (TAP) customizes perturbation strategies based on node attribute types, significantly strengthening unlearnability.
- MUGEN demonstrates effectiveness across various benchmarks and GNN backends, ensuring robustness against adversarial training.
Read more
MUGEN: Generating Unlearnable Graph Examples for Multiple Learning Tasks
Summary
The paper introduces MUGEN, a novel framework designed to generate unlearnable graph examples that protect against unauthorized representation learning across multiple tasks, including node classification, graph classification, and link prediction. Traditional methods typically focus on a single task, leaving data vulnerable to other potential uses. MUGEN addresses this limitation by producing a single perturbed dataset from a clean graph that maintains its topology while ensuring that models trained on it fail to generalize to clean data. The framework employs a shared Graph Neural Network (GNN) encoder with task-specific heads and incorporates a Task-Aligned Separability Objective (TASO) to enhance classwise separability and unlearnability. Additionally, it introduces Type-Adaptive Perturbation (TAP), which optimizes perturbation strategies based on the type of node attributes, improving the effectiveness of unlearnable examples. Experimental results demonstrate that MUGEN successfully generates transferable unlearnable graph examples across various GNN architectures and learning paradigms, maintaining effectiveness even under adversarial conditions.
Methodology
MUGEN utilizes a shared GNN encoder to generate a perturbed dataset that protects multiple tasks simultaneously. It employs a Task-Aligned Separability Objective (TASO) to optimize perturbations by incorporating classwise separability signals, and Type-Adaptive Perturbation (TAP) to tailor perturbation strategies based on the nature of node attributes, allowing for efficient optimization of both categorical and continuous features.
Results
Experiments conducted across five benchmarks and four GNN architectures demonstrate that MUGEN produces transferable unlearnable examples that effectively hinder unauthorized learning across node classification, graph classification, and link prediction tasks. The results indicate that the perturbations remain effective even under various adversarial conditions, including random feature masking and stochastic edge dropping.
Implications
The development of MUGEN has significant implications for data privacy and security in graph-based applications. By enabling the release of graph data that is robust against unauthorized learning, it can enhance the protection of sensitive information in domains such as social networks, molecular discovery, and scientific research. This framework can also inform future research on data protection strategies in machine learning.
QTEA: Ternary LLMs with Sparse Residual Salient Weight and By-Column Optimization
Large Language Models
Optimization
Efficient ML
- QTEA quantizes weights into a ternary base while using salient weights as residual compensators.
- The framework incorporates column-wise rescale refinement to enhance quantization expressiveness.
- An error decay mechanism is introduced to balance compensation strength during quantization.
- QTEA achieves a 16.7% accuracy improvement over the best existing ternary PTQ method.
Read more
QTEA: Ternary LLMs with Sparse Residual Salient Weight and By-Column Optimization
Summary
The paper introduces QTEA, a novel post-training quantization (PTQ) framework designed to optimize large language models (LLMs) by quantizing weights into ternary values. Traditional PTQ methods struggle with accuracy loss when reducing precision below 2 bits, particularly when using unstructured sparsity. QTEA addresses these challenges by employing salient weights as residual error compensators and implementing a semi-structured 1:4 sparsity scheme. The framework enhances hardware efficiency while maintaining model accuracy. Key innovations include a column-wise rescale refinement approach to improve ternary quantization and an error decay mechanism to mitigate order-dependent error propagation. Experimental results demonstrate that QTEA significantly outperforms existing ternary PTQ methods, achieving a 16.7% accuracy improvement on the Qwen3-14B model and notable reductions in perplexity on benchmark datasets. Additionally, the implementation of efficient CUDA kernels allows for substantial speedups in per-token generation, making QTEA a promising solution for deploying LLMs at scale.
Methodology
QTEA employs a ternary quantization approach that allocates residuals to salient weights using a semi-structured sparsity model. It refines column-wise scales during quantization and introduces an error decay mechanism to manage error propagation effectively. The framework is implemented with optimized CUDA kernels for efficient computation.
Results
On the Qwen3-14B model, QTEA compresses weights to an effective 1.7 bits per weight, improving average accuracy by 16.7% and achieving lower perplexity on WikiText and C4 datasets. Similar improvements are observed on the Llama3-8B model, with a 6.6% accuracy gain and significant reductions in perplexity. The implementation of CUDA kernels results in a 7.2× speedup in generation time over FP16.
Implications
QTEA's advancements in quantization techniques can facilitate the deployment of large language models in resource-constrained environments, making them more accessible for real-time applications and reducing operational costs.
MemoryWalker: Stop Training Agents on Contexts They Never Saw
Reinforcement Learning
Large Language Models
Theory
- Introduces the conditioning inconsistency problem in context-compressed training for agents.
- Proposes two exact solutions (LogitTree and packed 4D attention mask) for maintaining training consistency.
- Develops SDCC, a more efficient method that requires only a single backward pass and works with black-box systems.
- Demonstrates significant improvements in training consistency and performance metrics over naive methods.
Read more
MemoryWalker: Stop Training Agents on Contexts They Never Saw
Summary
The paper addresses the challenges faced by production agent harnesses that compress context during rollout, leading to conditioning inconsistencies in training. The authors identify that traditional methods of training on compressed contexts either cause 'time-travel leakage' or induce a train-inference mismatch. To resolve these issues, they propose two exact conditioning-level corrections: LogitTree and a packed 4D attention mask, both of which ensure gradient equivalence and maintain training consistency. However, these methods come with computational costs. To mitigate this, the authors introduce Self-Distillation for Conditioning Consistency (SDCC), which minimizes the KL divergence between the compressed student policy and a stop-gradient teacher. This approach allows for a single backward pass and can be applied to black-box harnesses. The evaluation of their framework across various context editors and black-box harnesses shows that naive compressed-stream training leads to significant log-probability gaps, while the proposed methods effectively close this gap, resulting in improved conditioning consistency, reduced logit drift, and higher rollout rewards.
Methodology
The authors formalize the context compression training as a conditioning inconsistency problem and propose two exact methods for traversing the conditioning tree. They also introduce SDCC, which minimizes KL divergence to align the behavior of the compressed policy with the original context, requiring only a single backward pass.
Results
The evaluation reveals that naive compressed-stream training inflates the train-rollout log-probability gap, particularly in eviction-heavy batches. In contrast, both LogitTree and the 4D attention mask maintain performance at the no-compression floor, while SDCC significantly reduces the gap, leading to lower logit drift and higher rollout rewards.
Implications
The findings suggest that addressing conditioning inconsistencies can enhance the performance of long-horizon agents in practical applications, particularly in environments where context compression is necessary for efficiency.
Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration
Computer Vision
- Introduces CORD, a method that preserves the original top-1 prediction during post-hoc calibration.
- Defines Top-1 Prediction Change Rate (TPCR) as a metric to measure the frequency of prediction changes.
- Demonstrates that CORD achieves zero TPCR while improving calibration metrics across multiple datasets.
- Eliminates the need for additional supervised fitting or hyperparameter tuning.
Read more
Let Confidence Change, Not the Prediction: Prediction-Preserving Repair for Post-hoc Calibration
Summary
This paper addresses the issue of post-hoc calibration in multiclass classification, where confidence scores can alter the top-1 prediction of a classifier. The authors introduce a novel method called Calibrator-Output Repair for Top-1 Decision Preservation (CORD), which ensures that the original top-1 prediction remains unchanged while allowing the confidence scores to be adjusted. CORD operates by repairing the full calibrated probability vector without modifying the fitted calibrator or requiring additional supervised fitting. The method calculates the mass assigned to the original top-1 prediction and redistributes the remaining mass among other classes, ensuring that the argmax of the repaired vector matches the original prediction. The authors demonstrate that CORD achieves zero Top-1 Prediction Change Rate (TPCR) across various datasets, including CIFAR-10/100 and ImageNet-1K, while also improving metrics such as Expected Calibration Error (ECE), Negative Log-Likelihood (NLL), and Brier score compared to direct outputs. This approach allows for effective calibration without sacrificing the integrity of the original predictions.
Methodology
CORD operates by analyzing the original and calibrated probability outputs to construct a repaired probability vector that retains the original top-1 prediction. It redistributes the probability mass while preserving the calibrated conditional distribution among other classes. The method does not alter the fitted calibrator or its outputs and does not require additional training or hyperparameter tuning.
Results
CORD achieves zero TPCR by design, indicating that the original top-1 predictions remain unchanged. Additionally, it lowers mean ECE, NLL, and Brier scores across CIFAR-10/100 and ImageNet-1K datasets compared to direct outputs, with improvements maintained under distribution shifts and across different calibration set sizes.
Implications
The findings suggest that CORD can be effectively used in practical applications where maintaining original predictions is critical, such as in safety-critical systems or when interpretability of model decisions is necessary. This method enhances the reliability of confidence scores without compromising the accuracy of predictions.