9 papers
Mismatch Matters: On-Policy Distillation Beyond Token Agreement
Zichao Yu, Chengzhi Yu, Shengze Xu +4
On-policy distillation (OPD) has emerged as a core component of modern LLM post-training pipelines, yet we reveal a failure mode: degenerate agreement, where students exploit repet…
ECHO-2: A Large-Scale Distributed Rollout Framework for Cost-Efficient Reinforcement Learning
Jingwei Song, Meng Chen, Jie Xiao +15
Reinforcement learning (RL) is a critical stage in post-training large language models (LLMs), involving repeated interaction between rollout generation, reward evaluation, and cen…
DLM-Scope: Mechanistic Interpretability of Diffusion Language Models via Sparse Autoencoders
Xu Wang, Bingqing Jiang, Yu Wan +3
Sparse autoencoders (SAEs) have become a standard tool for mechanistic interpretability in autoregressive large language models (LLMs), enabling researchers to extract sparse, huma…
SCULPT: Constraint-Guided Pruned MCTS that Carves Efficient Paths for Mathematical Reasoning
Qitong Fang, Haotian Li, Xu Wang
Automated agent workflows can enhance the problem-solving ability of large language models (LLMs), but common search strategies rely on stochastic exploration and often traverse im…
Does higher interpretability imply better utility? A Pairwise Analysis on Sparse Autoencoders
Xu Wang, Yan Hu, Benyou Wang +1
Sparse Autoencoders (SAEs) are widely used to steer large language models (LLMs), based on the assumption that their interpretable features naturally enable effective model behavio…
Feature Extraction and Steering for Enhanced Chain-of-Thought Reasoning in Language Models
Zihao Li, Xu Wang, Yuzhe Yang +3
Large Language Models (LLMs) demonstrate the ability to solve reasoning and mathematical problems using the Chain-of-Thought (CoT) technique. Expanding CoT length, as seen in model…