Ax Nicholas Potteiger, Ankita Samaddar, Taylor T. Johnson, Xenofon Koutsoukos 5/8/2026

Reward Shaping and Action Masking for Compositional Tasks using Behavior Trees and LLMs

Combines behavior trees with LLMs for automated reward shaping and action masking in compositional RL tasks, improving subtask learning efficiency.

Ax Zhiyuan Zhai, Xin Wang 5/8/2026

Selective Rollout: Mid-Trajectory Termination for Multi-Sample Agent RL

Introduces Selective Rollout to reduce training costs in GRPO by early-terminating rollouts with zero reward variance in multi-turn dialogue agent training.

Ax Elad Hoffer, Yochai Blau, Ron Banner, Daniel Soudry, Boris Ginsburg 5/8/2026

Retrieval from Within: An Intrinsic Capability of Attention-Based Models

Proposes INTRA framework enabling attention-based models to perform retrieval-augmented generation using only internal representations without external systems.

Ax Zegu Zhang, Jianhua Peng, Jian Zhang 5/8/2026

A Testable Certificate for Constant Collapse in Teacher-Guided VAEs

Theoretical analysis of constant collapse in variational autoencoders using teacher-guided training, establishing exact thresholds for this failure mode.

Ax Shen Xu, Xiangwen Zhuge, Zhe Xu, Yingkun Hu, Zheng Yang, Yunhao Liu 5/8/2026

HCInfer: An Efficient Inference System via Error Compensation for Resource-Constrained Devices

Introduces HCInfer, an inference system using error compensation to enable efficient LLM deployment on memory-constrained devices without significant accuracy loss.

Ax Yulong Huang, Xiang Liu, Hongxiang Huang, Xiaopeng Lin, Zunchang Liu, Xiaowen Chu, Zeke Xie, Bojun Cheng 5/8/2026

MDN: Parallelizing Stepwise Momentum for Delta Linear Attention

Proposes MDN for parallelizing momentum-based optimization in linear attention models to improve convergence and information retention in long-sequence LLM scaling.

Ax Hua-Dong Xiong 5/8/2026

Hypothesis generation and updating in large language models

Studies how LLMs generate and update hypotheses from partial information using the number game task, analyzing inference optimality and mechanisms.

Ax Samuel Blad, Martin L\"angkvist, Amy Loutfi 5/8/2026

Measuring Learning Progress via Gradient-Momentum Coupling

Proposes Gradient-Momentum Coupling metric to measure learning progress in curiosity-driven RL exploration.

Ax Carlo Romeo, Girolamo Macaluso, Alessandro Sestini, Andrew D. Bagdanov 5/8/2026

SOPE: Stabilizing Off-Policy Evaluation for Online RL with Prior Data

Proposes SOPE to stabilize off-policy evaluation in online reinforcement learning with prior data without manual tuning.

Ax Junhao Cai, Dohun Kim, Dowon Kim, Sung Il Choi, Chengjun Jin, Juhyun Park, Changhee Joo 5/8/2026

Retain-Neutral Surrogates for Min-Max Unlearning

Addresses machine unlearning via retain-neutral surrogates in min-max formulation to remove training data influence.

Ax Mohamed Lakas, Mohamed Amine Ferrag 5/8/2026

VARS-FL: Validation-Aligned Client Selection for Non-IID Federated Learning in IoT Systems

Proposes VARS-FL for client selection in federated learning on non-IID data using validation-aligned scoring.

Ax Cheng Xu, Xiaofeng Hou, Jiacheng Liu, Chao Li 5/8/2026

VisMMOE: Exploiting Visual-Expert Affinity for Efficient Visual-Language MoE Offloading

Presents VisMMoE system for efficient offloading of visual-language mixture-of-experts models on memory-constrained platforms.

Ax Miao Rang, Zhenni Bi, Hang Zhou, Kai Han, Xuechun Wang, An Xiao, Xinghao Chen, Yunhe Wang, Hanting Chen 5/8/2026

Near-Policy: Accelerating On-Policy Distillation via Asynchronous Generation and Selective Packing

Proposes Near-Policy Distillation to accelerate knowledge distillation for autoregressive models via asynchronous generation and selective packing.

Ax Zixuan Chen, Hao Lin, Zizhe Chen, Yizhou Tian, Garry Yang, Depeng Wang, Ya Guo, Huijia Zhu, James Cheng 5/8/2026

Knowing but Not Correcting: Routine Task Requests Suppress Factual Correction in LLMs

Studies failure mode where LLMs suppress factual corrections when embedded in task-oriented requests; benchmarks 8 models with suppression rates 19-90%.

Ax Eunseo Choi, Ho-Yeon Kim, Jaewon Lee, Taeyong jo, Myungjun lee, Heejin Ahn 5/8/2026

Uncertainty Estimation via Hyperspherical Confidence Mapping

Proposes Hyperspherical Confidence Mapping for uncertainty estimation in neural networks without sampling or distributional assumptions.

Ax Davide Maran, Csaba Szepesv\'ari 5/8/2026

Sharper Guarantees for Misspecified Kernelized Bandit Optimization

Theoretical analysis improving regret bounds for kernelized bandit optimization under model misspecification.

Ax Yoav Gelberg, Yam Eitan, Michael Bronstein, Yarin Gal, Haggai Maron 5/8/2026

Training Transformers for KV Cache Compressibility

Research on training transformers to be more compressible for KV cache compression, addressing long-context language modeling bottlenecks.

Ax Sicheng Ma, Tianyue Yang, Xiuzhe Wu, Xiao Xue 5/8/2026

Physical Fidelity Reconstruction via Improved Consistency-Distilled Flow Matching for Dynamical Systems

Paper on compressing high-fidelity flow-matching models for dynamical systems into efficient student networks for real-time inference.

Ax Yuntai Bao, Qinfeng Li, Xinyan Yu, Xuhong Zhang, Ge Su, Wenqi Zhang, Liu Yan, Haiqin Weng, Jianwei Yin 5/8/2026

Towards Steering without Sacrifice: Principled Training of Steering Vectors for Prompt-only Interventions

Research on training steering vectors for LLM behavior control without sacrificing generation quality or requiring per-vector tuning.

Ax Nobutaka Ono 5/8/2026

DiBA: Diagonal and Binary Matrix Approximation for Neural Network Weight Compression

DiBA proposes diagonal and binary matrix factorization for efficient neural network weight compression in linear layers and attention mechanisms.

Ax Aryeh Kontorovich, Kasper Green Larsen 5/8/2026

A Fine-Grained Understanding of Uniform Convergence for Halfspaces

Theoretical analysis of uniform convergence for halfspace learning, improving on standard VC dimension bounds.

Ax Ran Ben-Basat, William Kuszmaul, Michael Mitzenmacher, Amit Portnoy, Shay Vargaftik 5/8/2026

Quantizing With Randomized Hadamard Transforms: Efficient Heuristic Now Proven

Paper proving the efficiency of randomized Hadamard transforms for quantization in gradient compression, KV-cache compression, and vector search.

Ax Pei-Sen Li 5/8/2026

Matrix-Decoupled Concentration for Autoregressive Sequences: Dimension-Free Guarantees for Sparse Long-Context Rewards

Research paper presenting matrix-decoupled concentration bounds for autoregressive LLM sequences with improved dimension-free guarantees.

Ax Zheng Zhang, Cuong C. Nguyen, Kevin Wells, Gustavo Carneiro 5/8/2026

Multi-agent decision making: A Blackwell's informativeness approach

Principled approach analyzing multi-agent LLM decision-making using Blackwell informativeness theory with formal guarantees.

Ax Hugo Cazaux, Eyj\'olfur Ingi \'Asgeirsson, Hlynur Stef\'ansson 5/8/2026

Does Synthetic Data Help? Empirical Evidence from Deep Learning Time Series Forecasters

Large-scale empirical study on synthetic data augmentation for time series forecasting across architectures and datasets.

Ax Licheng Pan, Haochen Yang, Haoxuan Li, Yunsheng Lu, Yongqi Tong, Yinuo Wang, Shijian Wang, Zhixuan Chu, Lei Shen, Yuan Lu, Hao Wang 5/8/2026

Optimal Transport for LLM Reward Modeling from Noisy Preference

Framework using optimal transport theory to train robust reward models for RLHF despite noisy preference data.

Ax Saksham Rathi, Preeti, Mythili Vutukuru 5/8/2026

Requests of a Feather Must Flock Together: Batch Size vs. Prefix Homogeneity in LLM Inference

Studies trade-offs between batch size and prefix homogeneity in LLM inference to optimize memory-bound token generation.

Ax Duong Nguyen, Mohammed Jawhar, Nicolas Chesneau 5/8/2026

TFM-Retouche: A Lightweight Input-Space Adapter for Tabular Foundation Models

Input-space adapter for tabular foundation models enabling efficient task-specific adaptation without full fine-tuning.

Ax Yingxu Wang, Kunyu Zhang, Yanwu Yang, Thomas Wolfers, Yujie Wu, Siyang Gao, Nan Yin 5/8/2026

When Brain Networks Travel: Learning Beyond Site

Cross-site fMRI graph learning method addressing out-of-distribution generalization across brain imaging sites via transient neurodynamics.

Ax Mingcheng Zhu, Yu Liu, Tingting Zhu 5/8/2026

Towards Generation-Efficient Uncertainty Estimation in Large Language Models

Generation-efficient uncertainty estimation method for LLMs reducing inference cost of hallucination detection without full autoregressive generation.

Ax Kjetil Indrehus, Adrian Duric, Changkyu Choi, Ali Ramezani-Kebrya 5/8/2026

Towards Self-Explainable Document Visual Question Answering with Chain-of-Explanation Predictions

CoExVQA provides self-explainable Document Visual QA through chain-of-explanation predictions grounding answers to document evidence.

Ax Cristiano da Costa Cunha, Ajmal Mian, Tim French, Wei Liu 5/8/2026

Causal Reinforcement Learning for Complex Card Games: A Magic The Gathering Benchmark

MTG-Causal-RL benchmark for causal reinforcement learning on Magic: The Gathering with partial observation, masked action space, and structural causality.

Ax Maxim Fishman, Brian Chmiel, Ron Banner, Daniel Soudry, Boris Ginsburg 5/8/2026

Normalized Architectures are Natively 4-Bit

nGPT architecture with unit hypersphere constraints enables stable 4-bit LLM training without random transforms or per-tensor scaling.

Ax Trimble Chang, Yihang Liu, Mingjing Han, Han Zhang 5/8/2026

PRISM: Iterative Cross-Modal Posterior Refinement for Dynamic Text-Attributed Graphs

PRISM applies iterative cross-modal posterior refinement to dynamic text-attributed graph representation learning using multimodal fusion.

Ax Pierre Marion, Yu-Han Wu 5/8/2026

Understanding diffusion models requires rethinking (again) generalization

Position paper arguing diffusion model generalization requires fundamentally new theoretical frameworks beyond classical learning theory and benign overfitting.

Ax Mikalai Korbit, Mario Zanon 5/8/2026

Fast Gauss-Newton for Multiclass Cross-Entropy

Fast Gauss-Newton decomposes multiclass softmax cross-entropy curvature into true-vs-rest and within-competitor terms for scaled optimization.

Ax Yongyi Wang, Hanyu Liu, Lingfeng Li, Bozhou Chen, Ang Li, Qirui Zheng, Xionghui Yang, Chucai Wang, Wenxin Li 5/8/2026

Beyond Autoregressive RTG: Conditioning via Injection Outside Sequential Modeling in Decision Transformer

Extended Decision Transformer architecture conditioning method that injects Return-to-Go outside sequential modeling for improved offline RL efficiency.

Ax Yi-Siang Wang, Kuan-Yu Chen, Yu-Chen Den, Darby Tien-Hao Chang 5/8/2026

BoostLLM: Boosting-inspired LLM Fine-tuning for Few-shot Tabular Classification

BoostLLM applies boosting paradigm to LLM fine-tuning for few-shot tabular classification, improving performance in low-data regimes.

Ax Yun Qu, Qi Wang, Yixiu Mao, Heming Zou, Yuhang Jiang, Yingyue Li, Wutong Xu, Lizhou Cai, Weijie Liu, Clive Bai, Kai Yang, Yangkun Chen, Saiyong Yang, Xiangyang Ji 5/8/2026

Listwise Policy Optimization: Group-based RLVR as Target-Projection on the LLM Response Simplex

Listwise Policy Optimization reveals group-based RLVR methods for LLM post-training share common geometric structure as target-projection on response simplex.

Ax Samir Darouich, Vinh Tong, Llu\'is Pastor-P\'erez, Tanja Bien, Loay Mualem, Mathias Niepert 5/8/2026

SymDrift: One-Shot Generative Modeling under Symmetries

SymDrift enables one-shot generative modeling of physical systems via equivariant diffusion models that preserve global symmetries like rotations.

Ax Jiayi Zhao 5/8/2026

Matrix-Valued Optimism is Matrix-Valued Augmentation: Additive Hybrid Designs for Constrained Optimization

Mathematical proof extending equivalence between augmented Lagrangian and optimistic primal-dual methods to matrix-valued corrections for constrained optimization.

Ax Alireza Modirshanechi, Benjamin Eysenbach, Peter Dayan, Eric Schulz 5/8/2026

Unifying Goal-Conditioned RL and Unsupervised Skill Learning via Control-Maximization

Theoretical unification of goal-conditioned RL and unsupervised skill learning through control-maximization framework explaining why mutual information skill learning supports downstream goal-reaching.

Ax Yaomin Wang, Jianting Pan, Ran Tian, Xiaoyang Li, Yu Zhang, Hengle Qin, Tianshu YU 5/8/2026

AdaGamma: State-Dependent Discounting for Temporal Adaptation in Reinforcement Learning

AdaGamma proposes state-dependent discount factor method for deep reinforcement learning to improve planning horizon and bootstrapping stability in actor-critic algorithms.

Ax Xinrui Chen, Liu Yang, Ou Wu 5/8/2026

One Algorithm, Two Goals: Dual Scoring for Parameter and Data Selection in LLM Fine-Tuning

Proposes unified scoring mechanism for joint parameter and data selection in LLM fine-tuning to reduce computational overhead.

Ax Nicolas Samuel Blumer, Julien Martinelli, Samuel Kaski 5/8/2026

In-Context Black-Box Optimization with Unreliable Feedback

In-context black-box optimization handling unreliable side information from multiple feedback sources. Generalizes across feedback types.

Ax Dhruv Sarkar, Abhishek Sinha 5/8/2026

Constrained Contextual Bandits with Adversarial Contexts

Budget-constrained contextual bandit algorithm for adversarial contexts with cost constraints and realizability assumption.

Ax Mengfan Xu 5/8/2026

Bandit Learning in General Open Multi-agent Systems

Bandit learning framework for open multi-agent systems with dynamic agent arrivals/departures and endogenous non-stationarity.

Ax Muhammad Shahir Abdurrahman, Chun Deng, Azalia Mirhoseini, Philip Levis 5/8/2026

Federation of Experts: Communication Efficient Distributed Inference for Large Language Models

Federation of Experts architecture reduces communication bottlenecks in distributed MoE LLM inference by clustering experts per KV head.

Ax Xiaoyu Li, Andi Han, Jiaojiao Jiang, Junbin Gao 5/8/2026

Contrastive Identification and Generation in the Limit

Extension of identification-in-limit learning to contrastive settings with partially labeled data. Theoretical learning framework.

Ax Yi Wei, Xuan Qi, Furao shen, Jian Zhao, Vittorio Murino, Cigdem Beyan 5/8/2026

AffineLens: Capturing the Continuous Piecewise Affine Functions of Neural Networks

Framework for analyzing neural networks as continuous piecewise affine functions. Improves interpretability and expressivity analysis.