Ax Chongyu Fan, Pengfei Liu, Jingjia Huang, Sijia Liu, Yi Lin 22d ago

UP: Unbounded Positive Asymmetric Optimization for Breaking the Exploration-Stability Dilemma

Unbounded Positive optimization for RL addressing exploration-stability dilemma in LLM reasoning via importance sampling.

Ax Anthony Baez, Sheer Karny, Pat Pataranutaporn 22d ago

Dissociating the Internal Representations of Sycophancy in LLMs

Mechanistic analysis dissecting sycophancy in LLMs into factual and opinion subtypes using internal representations.

Ax Aoxiong Zeng, Yuxin Yang, Xiangquan Yang 22d ago

Online Data Selection Is Implicit Alignment

Study showing online data selection during fine-tuning acts as implicit alignment mechanism for LLM behavioral preferences.

Ax Meihua Dang, Stefano Ermon 22d ago

Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata

Constrained decoding for diffusion language models via finite automata enabling structured outputs like JSON schemas.

Ax Dennis Gross, Quentin Mazouni, Helge Spieker, Arnaud Gotlieb 22d ago

Gimitest: A Comprehensive Tool for Testing Reinforcement Learning Policies

Gimitest: Open-source comprehensive testing framework for single and multi-agent RL policies across varying conditions.

Ax Srikumar Krishnamoorthy 22d ago

Complexity-Budgeted, Interaction-Aware Interpretable Model for Tabular Data

Interpretable ML framework for tabular data addressing feature interactions through sparse rules and patterns.

Ax Alan Gany, Bogdan Cautis, Silviu Maniu 22d ago

Structural Adversarial Attacks on Relational Deep Learning under Integrity Constraints

Studies adversarial robustness of relational deep learning on heterogeneous temporal graphs with integrity constraints.

Ax Heye Huang, Jingguang Li, Zhiyuan Zhou, Paul Liang, Mingyu Wu, Kitae Jang, Jianqiang Wang 22d ago

A knowledge-augmented dataset of high-risk driving scenarios with LLM annotations for autonomous driving

K-Risk dataset combining high-risk driving scenarios with LLM annotations and semantic labels for autonomous driving safety research.

Ax Maximilian S. Ernst (Max Planck School of Cognition, Center for Lifespan Psychology Max Planck Institute for Human Development, Machine Learning Group Technische Universit\"at Berlin), Lorenz Linhardt (Machine Learning Group Technische Universit\"at Berlin, Berlin Institute for the Foundations of Learning and Data), Aaron Peikert (Center for Lifespan Psychology Max Planck Institute for Human Development), Oliver Eberle (Machine Learning Group Technische Universit\"at Berlin, Berlin Institute for the Foundations of Learning and Data) 22d ago

Distributed Sparse Interventions in Language Models

Studies causal interventions in language model components to understand task behavior, extending beyond global activation-space steering.

Ax Vladimir Gusev 22d ago

Fractal KV-Cache Archives: Lossless Symbolic Storage with In-Place Retrieval for Long-Context LLM Inference

Proposes lossless symbolic storage for KV-cache using contractive iterated-map codes to reduce memory cost in long-context LLM inference.

Ax Zhang Gongyue, Liu Donghan, Ren Weihong, Sheng Yixuan, Wang Zhiyong, Liu Honghai 22d ago

Information Allocation Dynamics in Neural Network Optimization

Analyzes optimizer implicit bias through information allocation dynamics, explaining how training signals distribute between weights.

Ax Zetian Hu, Shunyu Liu, Junjie Zhang, Yongcheng Jing, Ting-En Lin, Yongbin Li, Dacheng Tao 22d ago

Entropy Pacing Policy Optimization for Multi-Task Agentic Reinforcement Learning

Studies multi-task agentic RL for LLM-based agents, identifying exploration-exploitation dynamics across different tasks.

Ax Marcus Williams, Hannah Sheahan, Cameron Raymond, Tomek Korbak, Deng Pan, Peilin Yang, Leon Maksin, Ningyi Xie, Phillip Guo, Ian Kivlichan, Micah Carroll 22d ago

Predicting LLM Safety Before Release by Simulating Deployment

Method for pre-deployment safety evaluation of LLMs by simulating realistic deployments from de-identified conversations to assess failure rates.

Ax Zavier Li 22d ago

Geometric--Nongeometric Optimizer Calculus: A Modular Language for Reachable Gradient Methods

Modular language for characterizing reachable gradient methods, auditing optimizer mechanisms and their interactions systematically.

Ax Jinbo Yang, Mingyue Yuan, Boyuan Zhang, Yoshifumi Kitamura, Shikai Jing 22d ago

HPG-Diff: Hierarchical physics-guided diffusion with differentiable connectivity constraints for topology optimization

Proposes HPG-Diff, physics-guided diffusion framework for topology optimization with differentiable connectivity constraints.

Ax Georg Sch\"afer, Jakob Rehrl, Stefan Huber, Simon Hirlaender 22d ago

Safe Reinforcement Learning using Ideas from Model Predictive Control

Combines reinforcement learning with model predictive control to enforce hard safety constraints during exploration in cyber-physical systems.

Ax Donato Cerciello, Leonardo Schiavo, Angel Panizo-LLedot, Javier Huertas Tato, David Camacho 22d ago

FMMVCC: Fuzzy Mamba-based Multi-View Contrastive Clustering for Univariate Time Series

Presents FMMVCC, a Mamba-based clustering method for unsupervised time series analysis using fuzzy logic and contrastive learning.

Ax Chongkai Li, Bang Zhang, Wenjian Luo 22d ago

FedCVESA: Taking Away Training Data in Federated Learning via Correlation Value Encoding and Segmented Aggregation

Analyzes memorization-based privacy attacks (TATD) in federated learning, studying how malicious training can exfiltrate data in distributed settings.

Ax Pranav Sawant, Jakub Krej\v{c}\'i 22d ago

Mechanistic Interpretability for Neural Networks: Circuits, Sparse Features and Symbolic Reasoning

Comprehensive overview of mechanistic interpretability for reverse-engineering neural network internals, including circuit analysis and symbolic reasoning approaches.

Ax Zhiheng Zhou, Mengyao Zhou, Dengyi Zhao, Xingqin Qi, Guiying Yan 22d ago

Hypergraph Neural Stochastic Diffusion: An SDE Framework for Uncertainty Estimation

Proposes SDE framework for uncertainty estimation in hypergraph neural networks, addressing uncertainty from higher-order relations and complex dependencies.

Ax Oliver Makins, Orazio Angelini, Zohreh Shams, Mary Phuong 22d ago

Multi-Agent AI Control: Distributed Attacks Hamper Per-Instance Monitors

Studies multi-agent AI control techniques to prevent coordinated attacks across distributed AI deployments, addressing risks like model-weight exfiltration and training poisoning.

Ax Chethan Krishnamurthy Ramanaik, Tobias Callies, Michael Hecht, Eirini Ntoutsi 22d ago

On Adversarial Vulnerability of Vision-Language Models through the Lens of Intermediate Spectral Subspaces

Research on adversarial vulnerability in vision-language models through spectral analysis of intermediate linear transformations.

Ax Raiff H. Santos, Amilcar R. Queiroz, Tharcisyo S. S. Duarte, K. E. L. de Farias, Rafael A. Batista 22d ago

Generalist Vision-Language Models for Fast Radio Burst detection: a zero-shot benchmark against a specialized detector

Research benchmark evaluating open-weight vision-language models for fast radio burst detection; zero-shot generalist approach vs specialized detectors.

Ax Lo\"ic Cabannes, Pierre-Emmanuel Mazar\'e, Gergely Szilvasy, Matthijs Douze, Maria Lomeli, Ilze Amanda Auzina, Justin Carpentier, Gabriel Synnaeve, Herv\'e J\'egou 22d ago

Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity

Research introducing Sparse Delta Memory architecture for scaling linear RNNs; improves long-context recall while reducing FLOPs.

Ax Zhiyuan Li 22d ago

The Optimal Sample Complexity of Learning Autoregressive Chain-of-Thought

Theoretical analysis of sample complexity for learning autoregressive chain-of-thought traces; proves bounds governed by local next-token classification.

Ax Bojie Li, Noah Shi 22d ago

RLVP: Penalize the Path, Reward the Outcome

Research on reinforcement learning for real-world agents with irreversible interactions; proposes penalizing unsafe paths while rewarding outcomes.

Ax Vin\'icius Gabriel Angelozzi, H\'eber H. Arcolezi 22d ago

Where to Intervene? Benchmarking Fairness-Aware Learning on Differentially Private Synthetic Tabular Data

Research benchmark studying interaction between differential privacy and fairness-aware learning on synthetic tabular data in high-stakes ML.

Ax Athanasios Zeris 22d ago

FourierQK: Spectral Preprocessing of Query-Key Projections Improves Transformer Attention

Research: FFT-based spectral preprocessing of query-key projections improves transformer attention; 79% validation loss reduction on TinyShakespeare.

Ax Florian Fuchs, Jessy Gosselin-Grant, Boris Skuin, Michele Petteni, Alessandro Sestini, Joakim Bergdahl, Amir Baghi, Linus Gissl\'en 22d ago

Reward-Adaptive Iterative Discovery: A Case Study on Automated Game Testing for NHL26

RAID framework uses reward-adaptive RL for automated game testing, reducing retesting effort after behavior modifications in NHL26 development via iterative goalie AI exploit discovery.

Ax Jaris K\"uken, Shi Bin Hoo, Martin Mr\'az, Frank Hutter, Lennart Purucker 22d ago

TimEE: End-to-end Time Series Classification via In-Context Learning

TimEE applies in-context learning to end-to-end time series classification, replacing two-stage train-then-classify pipeline with unified learning from label demonstrations.

Ax Zhenyu Hou, Yujiang Li, Jie Tang, Yuxiao Dong 22d ago

Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning

Asynchronous single-rollout RL system for LLM post-training on long-horizon agentic tasks, improving efficiency and training stability over synchronous batch-interleaved approaches.

Ax Adam M. Oberman 22d ago

Fast Rates for Semi-Supervised Learning via Data-Augmentation Graph Regularization

Theoretical explanation for self-supervised learning efficiency: data augmentation induces graph structure enabling fast transductive rates O(1/n_L) with graph-Laplacian regularization.

Ax Maximilian Andreas Hoefler, Karsten Mueller, Wojciech Samek 22d ago

Collaborative Synthetic Data Generation for Knowledge Transfer in Federated Learning

Collaborative synthetic data generation for one-shot federated learning, enabling knowledge transfer across divergent client distributions in single communication round.

Ax Ana Schwengber Kelm, Christian Bockermann, J\"org Frochte 22d ago

Multi-Class vs. Multi-Label BERT for CVE-to-CWE Mapping: How Taxonomy Structure Shapes the Errors

Compares multi-class vs multi-label BERT formulations for CVE-to-CWE vulnerability mapping, analyzing how taxonomy structure shapes classification errors.

Ax C G Krishnanunni, Thomas Scott, Tan Bui-Thanh 22d ago

An optimal control approach for neural network architecture adaptation with a posteriori error estimation

Formulates neural network depth adaptation as optimal control problem with posteriori error estimation, enabling principled layer insertion based on approximation error distribution.

Ax Shiheng Zhang 22d ago

Guidance Breaks the Fitted Operator: A Terminal-Fitted Repair for Classifier-Free Guidance

Analyzes classifier-free guidance breakdown in diffusion models through numerical analysis, proposing terminal-fitted repair to stabilize sampling at high guidance scales.

Ax Xinyi Wu, Siyuan Liu, Ali Jadbabaie 22d ago

How Data Shapes RoPE Frequency Usage: From Positional Scale Matching to Length Generalization

Analyzes how RoPE frequency usage in transformers matches training data's relative-distance structure, with implications for length generalization and position embeddings.

Ax Xiangming Huang, Guannan Zhang, Lu Lu, Rapha\"el Pestourie 22d ago

Neural Operator-enabled Topology-informed Evolutionary Strategy for PDE-Constrained Optimization

NOTES integrates neural operators with evolutionary strategies for PDE-constrained inverse design optimization, combining robustness with high-dimensional search capability.

Ax Eric Zhu, Abhinav Shrivastava, Soumik Mukhopadhyay 22d ago

Selective Timestep Weighting and Advantage-Based Replay for Sample-Efficient Diffusion RLHF

Novel method combining selective timestep weighting and advantage-based replay to improve sample efficiency of RLHF applied to diffusion models, reducing feedback requirements.

Ax Lucas Pinto 22d ago

Calibration-Family Overfit: Why Trusted Sabotage Monitors Don't Transfer Across Lineages

Shows trusted monitoring models overfit to specific untrusted policy families and don't generalize across model lineages.

Ax Ya Wang, Hanwei Fan, Zhenguo Liu, Xiaofeng Zhou, Yangdi Lyu, Jiang Xu, Wei Zhang 22d ago

HiFuzz: Hierarchical Reinforcement Learning for Semantic-Aware and Adaptive CPU Fuzzing

HiFuzz uses hierarchical reinforcement learning with structured generation agents for CPU fuzzing and processor verification.

Ax Andrey Podivilov, Vadim Lomshakov, Sergey Savin, Matvei Startsev, Roman Pozharskiy, Maksim Parshin, Sergey Nikolenko 22d ago

AgentLens: Production-Assessed Trajectory Reviews for Coding Agent Evaluation

AgentLens benchmark evaluates coding agents on full trajectory quality including instruction-following, tool-use, and error recovery, not just pass/fail.

Ax Yu Cheng, Siyue Yao, Zhongang Qi, Shanyan Guan, Wei Li, Fajie Yuan 22d ago

Dynamic-in-Few-Step: Unifying Dynamic Computation and Few-Step Distillation for Efficient Video Generation

Dynamic-in-Few-Step combines dynamic computation with distillation to accelerate video diffusion model inference.

Ax Amin Haeri, Mahdi Ghelichi 22d ago

Specification Grounding Drives Test Effectiveness for LLM Code

Analysis of how specification-grounded tests improve LLM-generated code quality on edge cases compared to model self-testing.

Ax Felix Feldman, Joshua Harris, Timothy Laurence, Leo Loman, Ollie Higgins, Fan Grayson, Poonam Soma, Bethany Pace-Bonello, Michael Borowitz, Toby Nonnenmacher 22d ago

Healthier LLMs: Retrieval-Augmented Generation for Public Health Question Answering

Retrieval-Augmented Generation applied to medical QA reduces LLM hallucinations and grounds answers in current public health data.

Ax Issam Seddik, Sami Souihi, Mohamed Tamaazousti, Sara Tucci Piergiovanni 22d ago

The Power of Backdoor Absorption in Community Training

Method to detect and mitigate backdoor attacks in decentralized model training without full recomputation overhead.

Ax Zhangheng LI, Jianing Zhu, Junyuan Hong, Sungmin Eum, Shuowen Hu, Suya You, Zhangyang Wang 22d ago

POPS: Recovering Unlearned Multi-Modality Knowledge in MLLMs with Prompt-Optimized Parameter Shaking

POPS method recovers knowledge in multimodal LLMs after machine unlearning, addressing privacy-utility tradeoffs in MLLMs trained on sensitive data.

Ax Zeyuan Ding, Wenhai Liu, Yang Xu, Jiayu Hu, Yinda Chen, Yi Zhang, Yong Dai, Jian Tang, Xiaozhu Ju 22d ago

Pelican-VLA 0.5: Attending Before Acting Benefits Generalization

Vision-language-action model integrating perception, future prediction, and action planning with attention-based generalization without task-specific tuning.

Ax Inkyu Sa, Chanoh Park, Hea-Min Lee, Donghee Noh, Ho Seok Ahn 22d ago

Vision Language Action (VLA) Models for Unmanned Aerial Robotics and Bimanual Manipulation: A Review

Review of Vision Language Action models for embodied AI in robotics and manipulation tasks from camera images.

Ax Zavier Li 22d ago

Optimization Geometrodynamics: A Framework for Dynamic Geometric Optimization

Theoretical framework unifying gradient-based optimization as coupled evolution of parameters, particles, and time-varying Riemannian metrics.