InternReviewer & InternAdvocate: Objective Reward and Evaluation for Agentic Reinforcement Learning in Peer Review and Rebuttal
TLDR
Presents InternReviewer and InternAdvocate, scholarly agents for peer review and rebuttal trained via agentic RL with retrieval and citation verification to improve reasoning and accuracy.
Reasoning
The paper introduces a clear framework with a large dataset, retrieval tool, and unified reward system, and reports experimental gains in reasoning and citation accuracy. Its weakness is the narrow scope on peer review and rebuttal, limiting relevance to broader automated discovery or experimentation claims.
Read-first score
Read-first score 51.8, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 33.
Field roles
Rank sensitivity
Stability: volatile; rank range: 141.
Keyword Scores
Deep Analysis
Innovations
- Comprehensive framework for developing and evaluating specialized scholarly agents, InternReviewer and InternAdvocate
- Large-scale, high-quality scholarly dataset with an integrated high-efficiency arXiv retrieval tool for active evidence gathering
- Agentic reinforcement learning paradigm driven by a unified objective metric and reward system
- Objective multi-dimensional evaluation criteria including reference-anchored semantic alignment, structural compliance, and strict citation verification against real-time interaction logs to reduce hallucinations
Methodology
The work constructs a large-scale scholarly dataset and integrates an arXiv retrieval tool to support evidence gathering. InternReviewer and InternAdvocate are optimized using an agentic reinforcement learning loop with a unified objective reward based on reference-anchored semantic alignment, structural compliance, and citation verification against interaction logs. Evaluation focuses on reasoning depth and citation accuracy.
Key Results
Agents trained within the closed-loop framework show significant improvements in reasoning depth and citation accuracy.