RewardHarness: Self-Evolving Agentic Post-Training

🌐 Project | 💻 Code

Yuxuan Zhang*, Penghui Du*, Bo Li*, Cong Wei*, Junwen Miao, Huaisong Zhang, Songcheng Cai, Yubo Wang, Dongfu Jiang, Yuyu Zhang, Ping Nie, Wenhu Chen, Changqian Yu, Kelsey R. Allen

• RewardHarness introduces a self-evolving pipeline that turns execution feedback and automatic verifiers into scalable reward signals for agentic post-training, cutting the reliance on expensive human preferences.

SWE-QA-Pro: A Representative Benchmark and Scalable Training Recipe for Repository-Level Code Understanding

🤗 HF Collection | 💻 Code

Songcheng Cai*, Zhiheng Lyu*, Yuansheng Ni, Xiangchao Chen, Baichuan Zhou, Shenzhe Zhu, Yi Lu, Haozhe Wang, Chi Ruan, Benjamin Schneider, Weixu Zhang, Xiang Li, Andy Zheng, Yuyu Zhang, Ping Nie, Wenhu Chen

• SWE-QA-Pro introduces a repository-level QA benchmark and agentic training framework for code understanding, enabling LLM to navigate, reason over, and verify real-world software repositories.

VisCoder2: Building Multi-Language Visualization Coding Agents

🌐 Project | 🤗 HF Collection | 💻 Code

Yuansheng Ni*, Songcheng Cai*, Xiangchao Chen*, Jiarong Liang, Zhiheng Lyu, Jiaqi Deng, Kai Zou, Ping Nie, Fei Yuan, Xiang Yue, Wenhu Chen

• VisCoder2 introduces a multi-language visualization coding benchmark, dataset, and model family, enabling language models to generate, execute, render, and iteratively debug visualization code across diverse programming languages.

Revisiting Automated Evaluation for Long-form Table Question Answering

💻 Code

Yuqi Wang, Lyuhao Chen, Songcheng Cai, Zhijian Xu, Yilun Zhao

• LfTQA-Eval is a meta-evaluation dataset containing 2,988 human-annotated examples, designed to evaluate the performance of current automated metrics in assessing LLM-based LFTQA systems, with an emphasis on faithfulness and comprehensiveness.