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
DualEval: Joint Model-Item Calibration for Unified LLM Evaluation
Aaron J. Li, Hao Huang, Youngmin Park, Yitong Ma, Wei-Lin Chiang, Li Chen, Cho-Jui Hsieh, Bin Yu, Ion Stoica
Large Language Models NLP Interpretability
  • DUALEVAL unifies static benchmark correctness and open-ended preference signals for LLM evaluation.
  • The framework jointly estimates model abilities and item properties, enhancing evaluation stability and interpretability.
  • Empirical results show balanced model rankings and robust performance across various domains.
  • DUALEVAL enables diagnostic applications like benchmark compression and anomaly detection.
Read more
WattLayer: Get Layers Right to Estimate Inference Energy of Neural Networks
Adrien Sardi, Marie-Line Alberi Morel, Sara Alouf, Frédéric Giroire, Joanna Moulierac
Efficient ML
  • Introduction of WattLayer, a task-independent layer-wise energy estimation model.
  • Evaluation on a dataset of over 100,000 layers from 295 architectures, achieving a median error of 19.6%.
  • Demonstration of zero-shot generalization to new tasks without retraining.
  • Development of a comprehensive dataset and rigorous experimental protocol for energy measurement.
Read more
How Width and Data Shape Generalization Scaling Laws in Quadratic Neural Networks
Julius Girardin, Emanuele Troiani, Yizhou Xu, Vittorio Erba, Florent Krzakala, Lenka Zdeborová
Theory Optimization
  • Introduces a novel framework for analyzing generalization scaling laws in quadratic neural networks.
  • Characterizes generalization error as a function of model width, sample size, and regularization in a finite-sample setting.
  • Identifies distinct scaling regimes and transitions that affect generalization performance.
  • Demonstrates the influence of data structure on generalization through power-law relationships.
Read more
Uncertainty quantification via conformal prediction in data assimilation
Catherine George, Alireza Javanmardi, Tijana Janjić, Eyke Hüllermeier
Theory Time Series Efficient ML
  • Conformal prediction provides statistically rigorous uncertainty quantification with guaranteed coverage.
  • The study evaluates three variants of CP in a controlled atmospheric model setting.
  • CP-derived uncertainty estimates are compared with traditional ensemble-based measures.
  • Results highlight the strengths and limitations of CP in the context of data assimilation.
Read more
Autoencoder Architectures for Athlete Performance Scoring from Wearable Telemetry
Mateusz Kubita, Jan Zubalewicz, Krzysztof Siwek
Interpretability
  • Introduces a unified evaluation framework for unsupervised athlete ranking.
  • Demonstrates the effectiveness of deep autoencoders in reducing dimensionality of performance data.
  • Establishes a composite criterion for model selection that incorporates both reconstruction accuracy and interpretability.
  • Identifies key performance indicators such as running pace and heart rate as dominant factors in the latent score.
Read more
Decision-Aligned Evaluation of Uncertainty Quantification
Annika Schneider, Tommy Rochussen, Joshua Stiller, Vincent Fortuin
Theory
  • Introduces decision-alignment as a criterion for evaluating UQ metrics.
  • Demonstrates that many traditional UQ metrics are misaligned with decision-making utilities.
  • Proposes prior-weighted utility metrics for better alignment with downstream decisions.
  • Shows through experiments that prior-weighted metrics outperform conventional metrics in aligning with decision utility.
Read more
What Survives When You Compress a Recursive Reasoner for the Edge?
Pearse Jim, Steven Kolawole, Opegbemi Matthias Busoye, Glory Bagai, Virginia Smith
Efficient ML
  • Aggressive compression preserves local prediction accuracy but severely impacts global reasoning accuracy.
  • The collapse in reasoning performance is architectural, affecting MLP-mixing recursion but not attention mechanisms.
  • Carry-trajectory fidelity serves as a label-free signal to predict reasoning damage and recovery.
  • A deployment recipe is proposed that allows for efficient model compression suitable for edge hardware.
Read more
What Was That Again? Certified Robustness for Automatic Speech Recognition
Andrew C. Cullen, Neil Marchant, Jiani Xie, Paul Montague, Benjamin I.P. Rubinstein
Audio & Speech
  • Introduces a dual-gate certification pipeline for ASR systems to enhance robustness against perturbations.
  • Achieves up to a 55% reduction in Word Error Rate (WER) across multiple ASR architectures.
  • Provides both atomic and structural guarantees for sequence certification without requiring sentence alignment.
  • Demonstrates significant improvements in recall and reduced correlation between confidence scores and WER.
Read more
Recovering Governing Equations from Solution Data: Identifiability Bounds for Linear and Nonlinear ODEs
Yang Pan, Helmut Bölcskei
Theory
  • Introduces Hausdorff distance as a metric for comparing differential equations.
  • Establishes identifiability bounds for a wide class of ODEs.
  • Quantifies sample complexity required for reliable recovery of governing equations.
  • Addresses theoretical gaps in the uniqueness and stability of ODE identification.
Read more
Boundary condition fidelity for bottom-hole pressure and CO2 plume prediction in geological carbon storage
Romal Ramadhan, Seyyed A. Hosseini, Larry W. Lake
Theory Optimization Time Series
  • Boundary condition fidelity is critical for accurate BHP and CO2 plume predictions in GCS.
  • Uniform treatments that ignore corner storage lead to substantial pressure errors and plume misrepresentation.
  • Corner-adjusted boundary conditions significantly enhance prediction accuracy.
  • The gradual modifier with transmissibility correction offers the best performance across different reservoir types.
Read more
When Does Quality-Aware Multimodal Fusion Matter? A Leakage-Safe Diagnostic for Decision-Level Dependence
Jaden Moon, Arvind Pillai, Andrew Campbell
Multimodal
  • Introduces a leakage-safe diagnostic to assess the influence of quality signals on multimodal predictions.
  • Finds that permuting quality scores does not significantly degrade model performance, indicating minimal reliance on these scores.
  • Demonstrates that quality-aware fusion is beneficial only when quality estimates accurately identify reliable modalities.
  • Highlights the importance of distinguishing between correlation and causation in multimodal systems.
Read more
Finding Stationary Points by Comparisons
Helin Wang, Chenyi Zhang, Xiwen Tao, Yexin Zhang, Tongyang Li
Optimization Theory
  • Developed an algorithm for finding ϵ-stationary points using a comparison oracle with eO(n²/ϵ¹.⁵) queries.
  • Introduced a quantum algorithm that finds ϵ-stationary points with eO(n/ϵ¹.⁵) queries.
  • Improved dependence on ϵ compared to previous methods while incurring a higher cost in terms of dimensionality.
  • Identified the limitations of the comparison oracle model in accessing gradient information.
Read more
A Comparison of Fusion Techniques for Multi-Modal Human Activity Recognition on the HARMES Dataset
Ahmed Mohamady, Robin Burchard, Kristof Van Laerhoven
Multimodal Time Series
  • Systematic comparison of seven fusion techniques for multi-modal HAR on a common dataset.
  • Gated Multi-modal Fusion achieved the highest macro F1-score of 0.82.
  • Identified the contribution of each modality to overall performance.
  • Demonstrated that multi-modal fusion can reduce the impact of handedness on recognition accuracy.
Read more
Training Observable Control Policies to Expose Agent State Through Actions
Andres Enriquez Fernandez, John J. Bird
Reinforcement Learning Robotics Theory
  • Introduces a method for estimating agent states using only observable actions.
  • Implements a reinforcement learning framework to enhance policy observability.
  • Demonstrates improved estimator performance with minimal impact on task performance.
  • Focuses on applications in autonomous agent coordination under communication limitations.
Read more
Difference of Convex Programming in the Wasserstein Space with Applications to MMD Optimization
Clément Bonet, Pierre-Cyril Aubin-Frankowski, Youssef Mroueh
Optimization Theory Generative Models
  • Introduction of the Wasserstein Convex-Concave Procedure (WCCCP) for optimizing non-convex functionals.
  • Theoretical guarantees of almost stationarity for the proposed optimization scheme.
  • Empirical results show faster and more stable convergence compared to Wasserstein gradient descent.
  • Focus on Maximum Mean Discrepancy (MMD) and Energy Distance (ED) functionals.
Read more
The Curse of Multiple Mediators: Hidden Interaction Effects in Activation Patching
Sankaran Vaidyanathan, David Arbour, Aaron Mueller, Scott Niekum, David Jensen
Interpretability Theory Large Language Models
  • Activation patching's natural indirect effect (NIE) includes hidden interaction effects (INT) that can misrepresent component importance.
  • INT varies with the distance between clean and patched activations and is negligible in locally affine models.
  • The presence of INT explains known failures in activation patching, particularly in the GPT-2 IOI circuit.
  • Ranking components solely by pure indirect effect (PIE) can lead to significant inaccuracies.
Read more
Zero-Shot Size Transfer for Neural ODEs on Sparse Random Graphs: Graphon Limits and Adjoint Convergence
Mingsong Yan, Zhida Wang, Sui Tang
Graph Learning Theory Efficient ML
  • Establishes a quantitative theory for zero-shot size transfer in GNDEs on sparse random graphs.
  • Proves trajectory-wise convergence of GNDE solutions to Graphon-NDE solutions with a specific convergence rate.
  • Demonstrates asymptotic consistency of DTO and OTD training paradigms for GNDEs.
  • Validates theoretical findings through experiments on various graphon classes.
Read more
TeRoR: Decoupled Temporal Rotation with Relational Circular Region for Temporal Knowledge Graph Embedding
Peijia Xie, Yike Liu, Chao He, Huiling Zhu
Graph Learning Time Series
  • TeRoR enhances temporal information representation by decoupling the temporal influence on subject and object entities.
  • The model introduces a relation-aware circular region to effectively capture complex multi-relational interactions.
  • Experimental results show significant performance improvements over existing state-of-the-art temporal knowledge graph embedding models.
  • TeRoR addresses limitations in existing models regarding the mapping properties of various relations.
Read more
Retroactive Advantage Correction: Closed-Form V-Trace Bias Correction for Delay-Aware RLHF
Arnav Raj
Reinforcement Learning Theory Optimization
  • Introduces Retroactive Advantage Correction (RAC) to handle delayed rewards in RLHF.
  • Proves that RAC provides an unbiased correction 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
Towards Automating Scientific Review with Google's Paper Assistant Tool
Rajesh Jayaram, Drew Tyler, David Woodruff, Corinna Cortes, Yossi Matias, Vahab Mirrokni, Vincent Cohen-Addad
Theory Large Language Models Efficient ML
  • Introduction of the Paper Assistant Tool (PAT) for automating scientific review.
  • PAT improves error detection in mathematical proofs by 34% over traditional methods.
  • Pilot deployments at major conferences showed positive community feedback.
  • A proposed taxonomy outlines levels of AI-human collaboration in peer review.
Read more
Disentangling Continuous-Time Latent Dynamics: Identifiability of Latent SDEs via Diffusion Shifts
Yuanyuan Wang, Wenjie Wang, Haoxuan Li, Mingming Gong, Kun Zhang
Time Series Theory
  • Identifiability of continuous-time latent SDEs is achieved using diffusion shifts.
  • Two diagonal diffusion regimes with distinct variance ratios can identify latent coordinates.
  • 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
EVOM: Agentic Meta-Evolution of Actor-Critic Architectures for Reinforcement Learning
Boyun Zhang, Chao Wang, Kai Wu
Reinforcement Learning Large Language Models Optimization
  • EVOM automates the design of actor-critic architectures in reinforcement learning.
  • The framework uses a bi-level optimization approach combining low-fidelity PPO and an LLM-based design agent.
  • Experimental results show significant performance improvements over traditional and state-of-the-art methods.
  • Ablation studies validate the importance of both the meta-evolution loop and the LLM design agent.
Read more
Physics-guided Convolutional Neural Network for Domain Growth Prediction in Systems with Conserved Kinetics
Vijay Yadav, Madhu Priya, Manish Dev Shrimali, Prabhat K. Jaiswal
Theory Efficient ML
  • Introduction of an attention-based, physics-guided CNN for modeling phase separation.
  • Incorporation of conservation constraints in the loss function to maintain order parameter consistency.
  • Demonstration of the model's ability to predict long-term dynamics accurately.
  • Validation of the model against known growth laws, confirming its physical relevance.
Read more
COCOLogic-V2: Identifying Logical Inconsistencies via Truly Hard-Negatives
David Steinmann, Antonia Wüst, Kristian Kersting, Wolfgang Stammer
Computer Vision Interpretability
  • Introduction of COCOLogic-V2, enhancing the scope of visual inductive reasoning tasks.
  • Dataset categorization into positive variants, near-boundary, and far-from-boundary negatives for better model evaluation.
  • Current interpretable models perform well on clear cases but fail on near-boundary samples, indicating a lack of true logical understanding.
  • COCOLogic-V2-FS provides a resource for few-shot learning in complex reasoning tasks.
Read more
Flexformer: Flexible Linear Transformer with Learnable Attention Kernel
Haoran Zhang, Feng Zhou
NLP Efficient ML Theory
  • Flexformer utilizes learnable attention kernels to achieve linear complexity in attention mechanisms.
  • The model can learn a wide variety of attention patterns, including softmax attention.
  • Flexformer demonstrates superior performance in language modeling and sequence classification tasks.
  • It shows strong kernel transferability across different domains.
Read more
RecallRisk-BERT: A Multi-Task Framework for Post-Report Medical Device Recall Triage
Ali Semih Atalay, Sevgi Yigit-Sert
NLP Multimodal
  • Introduces RecallRisk-BERT, a multi-task learning framework for medical device recall triage.
  • Utilizes a large dataset of FDA recall records to improve prediction accuracy for recall severity and root causes.
  • Demonstrates that joint modeling of severity and root-cause categories enhances performance compared to single-task models.
  • Achieves high accuracy and strong consistency with observed data, indicating the effectiveness of text–tabular learning.
Read more
Empirical Software Engineering TerraProbe: A Layered-Oracle Framework for Detecting Deceptive Fixes in LLM-Assisted Terraform
Manar Alsaid, Chimdumebi Nebolisa, Faris Abbas
Large Language Models
  • Introduction of TerraProbe, a five-layer oracle framework for evaluating LLM-generated Terraform repairs.
  • Demonstration that deceptive fixes are systemic across multiple LLMs, with rates ranging from 57.1% to 71.4%.
  • Development of a formal taxonomy for deceptive fixes, validated with high inter-rater reliability.
  • Statistical analysis revealing significant discrepancies between superficial success metrics and deeper evaluation criteria.
Read more
Deterministic Pareto-Optimal Policy Synthesis for Multi-Objective Reinforcement Learning
Aniruddha Joshi, Niklas Lauffer, Sanjit Seshia
Reinforcement Learning Optimization Theory
  • Introduction of a preference-conditioned Bellman operator for MOMDPs.
  • Proof of convergence to the Pareto-optimal values and coverage of the Pareto frontier.
  • Extraction of deterministic policies from converged Q-estimates with single-step transition memory.
  • Empirical validation demonstrating the algorithm's ability to recover complex trade-offs.
Read more
Localizing RL-Induced Tool Use to a Single Crosscoder Feature
Andrii Shportko, Shubham Bhokare, Ahmed Zeyad A Alzahrani, Bowen Cheng, Gustavo Mercier, Jessica Hullman
NLP Large Language Models Reinforcement Learning
  • Introduces Dedicated Feature Crosscoders (DFC) to isolate RL-specific features in language models.
  • Demonstrates a +31.1% improvement in tool correctness through encode-decode reconstruction.
  • Identifies capability spillover, achieving a +6.8% increase in tool-correctness in a frozen model.
  • Shows that steering a single A-exclusive neuron can lead to a +65.0% improvement in tool-correctness.
Read more
RSPC: A Benchmark for Modeling Stress and Psychiatric Conditions in Digitally Mediated Relationships using Psychiatrist Annotations
Parmitha Vangapandu, Sai Ganesh Mokkapati, Sathwik Narkedimilli, MSVPJ Sathvik, Timothy Liu, Simon See, Johannes C. Eichstaedt
NLP Large Language Models
  • Introduction of RSPC, the first benchmark linking psychiatric conditions with relational stressors in digital communication.
  • Utilization of psychiatrist annotations for a clinically grounded approach to mental health modeling.
  • Benchmarking of various transformer models and LLMs reveals distinct capabilities in handling relationally contextualized mental health tasks.
  • Strong associations found between anxiety disorders and relational uncertainties, emphasizing the importance of context in psychiatric inference.
Read more
HybridCodec: Modeling Discrete and Continuous Representations for Efficient Speech Language Models
Artem Ploujnikov, Francesco Verdini, Samir Sadok, Mirco Ravanelli
Audio & Speech Multimodal Large Language Models
  • HybridCodec combines discrete and continuous audio representations to mitigate information loss.
  • The architecture includes a hybrid Transformer that supports both autoregressive and non-autoregressive predictions.
  • Experimental results show significant improvements in speaker characteristic retention and reduced autoregressive steps.
  • The framework effectively handles multiple speech tasks within a single model.
Read more
CPAgents: Agentic Composite Phenotype Generation for Cardiac Disease Association
Zuoou Li, Wenlong Zhao, Kelly Yu, Weitong Zhang, Paul M. Matthews, Wenjia Bai, Bernhard Kainz, Mengyun Qiao
Interpretability
  • CPAgents automates the discovery of composite phenotypes, enhancing the ability to capture non-linear effects and interactions.
  • The framework consists of three agents: Analyst, Proposer, and Verifier, which work iteratively to construct and validate phenotypes.
  • Evaluation on a large cardiac imaging dataset showed that CPAgents outperformed baseline methods in disease discrimination.
  • The discovered phenotypes are interpretable and reproducible, providing clear evidence trails for clinical application.
Read more
fTNN: a tensor neural network for fractional PDEs
Qingkui Ma, Hehu Xie, Xiaobo Yin
Theory Optimization
  • Introduction of fTNN, a novel tensor neural network for fractional PDEs.
  • Development of a deterministic integration framework for the fractional Laplacian.
  • Construction of boundary-singularity-aware trial functions to enhance solution accuracy.
  • Design of a spatiotemporally separable neural network for time-dependent fractional PDEs.
Read more
Effective Covariance Dynamics in Solvable High-Dimensional GANs
Andrew Bond, Zafer Doğan
Generative Models Theory Optimization
  • Derivation of high-dimensional effective covariance dynamics for multi-feature GANs.
  • Identification of a spectral solvable region that governs learning stability and recovery.
  • Demonstration of a signal-boosting mechanism through low-rank correlations.
  • Empirical validation showing improved recovery of data-driven subspaces with informed generator covariance.
Read more
PEBS: Per-rater Empirical-Bayes Shrinkage for RLHF Reward-Model Calibration
Arnav Raj
Reinforcement Learning
  • PEBS provides a solution to the calibration issues in RLHF by allowing for per-rater adjustments.
  • The method utilizes empirical-Bayes shrinkage to improve the accuracy of individual annotator calibrations.
  • PEBS significantly reduces RMSE in reward model predictions compared to traditional pooled approaches.
  • The approach is validated on multiple datasets, demonstrating its robustness and applicability.
Read more
Dangerous Liaisons of Convex Learning and Non-Affine Aggregation
Thomas Boudou, Batiste Le Bars, Nirupam Gupta, Aurélien Bellet
Theory Optimization
  • Monotonicity of aggregated gradients is preserved only by positively affine aggregation rules.
  • Non-affine aggregation leads to failures in last-iterate convergence and increased instability in algorithms.
  • The paper provides a unified theoretical framework for understanding the limitations of non-affine aggregation in convex learning.
  • Sufficient conditions for restoring monotonicity in non-affine aggregation are identified.
Read more
OverFlowLight: Real-Time Gridlock Prevention and Traffic Signal Optimization for Urban Intersections
Mingyuan Li, Boyang Huang, Tianqi Jiang, Chenpu Li, Chunyu Liu, Yang Li, Ruimin Li, Qiang Wu
Reinforcement Learning Optimization Computer Vision
  • OverFlowLight effectively detects and mitigates traffic overflow in real-time.
  • The framework integrates multi-modal sensing for accurate overflow detection.
  • Dynamic overflow phases are generated to clear blocking queues, improving traffic flow.
  • Real-world deployments show significant reductions in overflow incidents and increased throughput.
Read more
Stochastic Gradient Optimization with Model-Assisted Sampling
Jonne Pohjankukka, Jukka Heikkonen
Optimization Efficient ML Theory
  • Introduces a model-assisted sampling framework to reduce variance in stochastic gradient estimation.
  • Bridges concepts from machine learning optimization and survey sampling theory.
  • Empirical results indicate performance gains in a majority of experiments, particularly for medium-sized datasets.
  • The proposed method integrates with existing optimizers, enhancing efficiency without altering their dynamics.
Read more
Escaping Iterative Parameter-Space Noise: Differentially Private Learning with a Hypernetwork
Naoki Nishikawa, Shokichi Takakura, Satoshi Hasegawa
Theory Efficient ML Generative Models
  • Introduces a hypernetwork-based framework for differentially private learning that reduces noise impact.
  • DP-DeepSets architecture generates model parameters from a low-dimensional dataset embedding with a single noise injection.
  • Theoretical analysis shows improved utility compared to traditional DP-SGD methods.
  • Demonstrates superior performance in LoRA fine-tuning of diffusion models using limited private data.
Read more
PersistentKV: Page-Aware Decode Scheduling for Long-Context LLM Serving on Commodity GPUs
Muhammad Ahmed
Large Language Models Efficient ML Optimization
  • PersistentKV improves long-context LLM serving by optimizing KV-cache management.
  • The system employs a native block-table decode engine to enhance efficiency.
  • An adaptive scheduling policy selects between PersistentKV and FlashInfer based on workload characteristics.
  • Workqueue scheduling significantly reduces launch fan-out, improving throughput.
Read more
Implementation of reinforcement learning in chemical reaction networks: application to phototaxis as curiosity-driven exploration
Ruyi Tang, Grégoire Sergeant-Perthuis, David Colliaux
Reinforcement Learning Robotics Theory
  • Introduces a framework combining POMDPs with CRN dynamics for modeling phototaxis.
  • Demonstrates that tumbling behavior in algae is an adaptive strategy for information acquisition.
  • Uses IRL to derive a phototactic policy from experimental trajectory data.
  • Establishes a connection between sensory geometry, subjective inference, and biochemical implementation.
Read more
Dual-Learning based Penalized Multi-Align Clustering for Multi-View Incomplete and Disorderly Data
Liang Zhao, Shubin Ma, Bo Xu, Qingchen Zhang
Multimodal
  • Introduces DLPMAC, a model for aligning and fusing incomplete multimodal data.
  • Utilizes Dual-Learning to maintain semantic and structural consistency across modalities.
  • Employs a penalty mechanism to improve alignment accuracy and prevent excessive sample aggregation.
  • Demonstrates effectiveness through experimental validation in real-world scenarios.
Read more
Halt Fast! Early Stopping for Certified Robustness
Andrew C. Cullen, Paul Montague, Benjamin I.P. Rubinstein
Theory Efficient ML Computer Vision
  • Introduces a meta-learning framework for anytime-valid certified robustness.
  • Achieves a 20-fold reduction in sample complexity compared to traditional RS methods.
  • Enables adaptive termination conditions based on application-specific risk thresholds.
  • Demonstrates potential for real-time applications in safety-critical environments.
Read more
GEOALIGN: Geometric Rollout Curation for Robust LLM Reinforcement Learning
Ting Zhou, Zhenqing Ling, Yiyang Zhao, Ying Shen, Daoyuan Chen
Reinforcement Learning Large Language Models
  • Identification of directional inconsistency as a significant failure mode in online RL for LLMs.
  • Introduction of GEOALIGN, a lightweight module for effective rollout curation.
  • Demonstrated improvements in performance and stability over existing robust RL methods.
  • GEOALIGN operates without requiring per-rollout policy gradients, ensuring efficiency.
Read more
Neural Architecture Search for Generative Adversarial Networks: A Comprehensive Review and Critical Analysis
Abrar Alotaibi, Moataz Ahmed
Generative Models Optimization
  • NAS significantly improves the design and performance of GANs.
  • Evolutionary algorithms and gradient-based methods show superior performance in certain contexts.
  • Robust evaluation metrics are essential for accurately assessing GAN performance.
  • Diverse datasets are crucial for evaluating the effectiveness of GAN architectures.
Read more
Explaining Temporal Graph Neural Networks via Feature-induced Information Flow
Ping Xiong, Thomas Schnake, Klaus-Robert Müller, Shinichi Nakajima
Graph Learning Interpretability Time Series
  • Introduction of a novel Event Relevance (ER) method that captures the entire information flow in ETGNNs.
  • Extension of the Normalized Relevance Measure (NRM) framework to facilitate modular decomposition for complex networks.
  • Demonstration of superior performance in qualitative and quantitative evaluations compared to existing explanation methods.
  • Ability to analyze higher-order interactions among events, enhancing the interpretability of model predictions.
Read more
PerturbCellRL: Verifier-Guided Reinforcement Learning for Single-Cell Perturbation Prediction
Dongxia Wu, Mingyu Li, Yuhui Zhang, Anurendra Kumar, Emma Lundberg, Serena Yeung-Levy, Emily B. Fox
Reinforcement Learning Generative Models
  • PerturbCellRL incorporates biological verifiers as reward functions to enhance single-cell perturbation predictions.
  • The framework improves individual cell response plausibility while maintaining population-level distributional quality.
  • Evaluation on multiple benchmarks shows significant improvements in reward-aligned metrics.
  • The approach emphasizes the importance of biological consistency in generative modeling for computational biology.
Read more
Learning to Reason with Curriculum II: Compositional Generalization
Nived Rajaraman, Audrey Huang, Miroslav Dudik, Robert Schapire, Dylan Foster, Akshay Krishnamurthy
Theory Reinforcement Learning Large Language Models
  • Compositional generalization is essential for effective reasoning in AI.
  • An autocurriculum approach significantly reduces the learning complexity compared to direct methods.
  • In supervised fine-tuning, the curriculum allows learning from only 2 e^O(√log T) tokens.
  • In reinforcement learning, the curriculum relaxes the coverage requirement on the reference model.
Read more