Reproduce “DeepResearch RewardBench: Evaluating Reward Models for Deep Research Report” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Shudong Liu, Yuting Zhong, Guanhua Chen, Xinyi Yang, Junnan Liu, Lidia S. Chao, Jiaheng Liu, Derek F. Wong
DeepResearch RewardBench tests how well models rank long research reports. It uses reports from research agents and search-enabled LLMs, ranking them with task-specific rubrics and human checks. Comparative judgment outperforms independent scoring in tested settings. Gemini-2.5-Flash scores 78.87% on the weighted average, versus 63.75% for the strongest sequence-classifier baseline, but trails it on quadruplet full ranking.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.