6 papers
Scaling Verifiable Environments for Long-horizon Work Agents
Jiazheng Zhang, Long Ma, Yunxian Yang +24
Work agents operate over digital artifacts to execute professional knowledge-intensive work, requiring training environments that support long-horizon interaction and trustworthy v…
A Token-Level Analysis of Sampled-Token Reverse-KL On-Policy Distillation
Bing Shao, Jiazheng Zhang, Long Ma +12
On-policy distillation (OPD) supervises a student on its own trajectories with token-level signals from a frozen teacher, yet how a sampled loss allocates updates across tokens rem…
Entropy Polarity in Reinforcement Fine-Tuning: Direction, Asymmetry, and Control
Jiazheng Zhang, Ziche Fu, Junrui Shen +17
Policy entropy has emerged as a fundamental measure for understanding and controlling exploration in reinforcement learning with verifiable rewards (RLVR) for LLMs. However, existi…
Thinking with Constructions: A Benchmark and Policy Optimization for Visual-Text Interleaved Geometric Reasoning
Haokun Zhao, Wanshi Xu, Haidong Yuan +3
Geometric reasoning inherently requires "thinking with constructions" -- the dynamic manipulation of visual aids to bridge the gap between problem conditions and solutions. However…
Steering LLMs via Scalable Interactive Oversight
Enyu Zhou, Zhiheng Xi, Long Ma +9
As Large Language Models increasingly automate complex, long-horizon tasks such as \emph{vibe coding}, a supervision gap has emerged. While models excel at execution, users often s…
TaxoBench: Can Deep Research Agents Retrieve and Organize? Evaluating the Synthesis Gap with Expert Taxonomies
Ming Zhang, Jiabao Zhuang, Wenqing Jing +19
Deep Research Agents increasingly automate survey writing, yet existing benchmarks do not jointly test whether they retrieve the papers experts consider essential and organize thos…