Tianyu Pang

Variational Reasoning for Language Models

Think in Parallel, Answer as One: Logit Averaging for Open-Ended Reasoning

Muon Outperforms Adam in Tail-End Associative Memory Learning

Reinforcing Query-Level Meta-Agents

Fostering Video Reasoning via Next-Event Prediction

Language Models Can Learn from Verbal Feedback Without Scalar Rewards

Verl-Tool: Towards Holistic Agentic Reinforcement Learning with Tool Use

Nonparametric Data Attribution for Diffusion Models

Reinforcing General Reasoning Without Verifiers

Anatomy of a Hybrid Mind: Deconstructing Hybrid Reasoning in Large Language Models

FreezeVLA: Action-Freezing Attacks against Vision-Language-Action Models

Imperceptible Jailbreaking against Large Language Models

Orient Anything V2: Unifying Orientation and Rotation Understanding

Understanding R1-Zero-Like Training: A Critical Perspective

Cheating Automatic LLM Benchmarks: Null Models Achieve High Win Rates

When Attention Sink Emerges in Language Models: An Empirical View

NoisyRollout: Reinforcing Visual Reasoning with Data Augmentation

RegMix: Data Mixture as Regression for Language Model Pre-training

Improving Your Model Ranking on Chatbot Arena by Vote Rigging

Adversarial Attacks against Closed-Source MLLMs via Feature Optimal Alignment

Efficient Process Reward Model Training via Active Learning

Scaling up Masked Diffusion Models on Text

Optimizing Anytime Reasoning via Budget Relative Policy Optimization

SkyLadder: Better and Faster Pretraining via Context Window Scheduling

Lifelong Safety Alignment for Language Models

Improved Techniques for Optimization-Based Jailbreaking on Large Language Models

Weak-to-Strong Jailbreaking on Large Language Models

A Closer Look at Machine Unlearning for Large Language Models

LLM-based Multi-Agents System Attack via Continuous Optimization with Discrete Efficient Search

Bootstrapping Language Models with DPO Implicit Rewards

Improving Long-Text Alignment for Text-to-Image Diffusion Models

Orient Anything: Learning Robust Object Orientation Estimation from Rendering 3D Models

Weak-to-Strong Jailbreaking on Large Language Models

SimLayerKV: A Simple Framework for Layer-Level KV Cache Reduction

Unnatural Languages Are Not Bugs but Features for LLMs

BanditSpec: Adaptive Speculative Decoding via Bandit Algorithms

Test-Time Backdoor Attacks on Multimodal Large Language Models

Meta-Unlearning on Diffusion Models: Preventing Relearning Unlearned Concepts

Denial-of-Service Poisoning Attacks against Large Language Models