Reproduce “Dr.Credit: Rubric-Grounded Process Credit Assignment for Deep Research Agents” 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
Yingjian Zhu, Zhenyi Wang, Jiaxin Guo, Kun Ding, Ying Wang, Shen Huang, Xunjie Zhu, Pengjun Xie and 1 more
Dr.Credit trains deep-research agents to reward tool turns that add support to report requirements, alongside evaluating the final report. For each rubric, an LLM judge compares a tool response with support already accepted in that trajectory, crediting new or partial evidence. In matched Qwen3-8B experiments, it scores 61.7 versus GRPO’s 58.2 averaged across four benchmarks; results cover deep research.
Key insights
- Evidence is judged against rubric-specific requirements rather than a canonical answer, allowing partial contributions to open-ended tasks to receive credit.
- Maintaining an evidence history helps distinguish newly useful information from material that merely repeats what the agent has already found.
- Process and outcome supervision contribute independently; the ablations also indicate that search navigation and snippet scoring can add value beyond webpage visits alone.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.