InternReviewer & InternAdvocate: Objective Reward and Evaluation for Agentic Reinforcement Learning in Peer Review and Rebuttal
The paper proposes RL-trained scholarly agents whose rewards verify citation use against retrieval logs.
InternReviewer and InternAdvocate are built for peer review and rebuttal writing, with arXiv retrieval included for evidence gathering. The framework uses a unified objective metric instead of subjective model judging. Its reward criteria cover semantic alignment to references, structural compliance, and citation checks meant to catch hallucinated support. The authors report improved reasoning depth and citation accuracy in experiments. HF Daily Papers' note
InternReviewer and InternAdvocate are built for peer review and rebuttal writing, with arXiv retrieval included for evidence gathering. The framework uses a unified objective metric instead of subjective model judging. Its reward criteria cover semantic alignment to references, structural compliance, and citation checks meant to catch hallucinated support. The authors report improved reasoning depth and citation accuracy in experiments. HF Daily Papers' note
score 4