AI Agents Objectively Evaluate Peer Reviews and Rebuttals

This paper introduces a new framework for developing and evaluating AI agents specialized in academic peer review and rebuttal processes. The system, named InternReviewer and InternAdvocate, aims to produce professional scholarly content by combining domain-specific reasoning with factual verification. A key innovation is an agentic Reinforcement Learning paradigm that uses a unified, objective reward system. This system incorporates multi-dimensional criteria such as semantic alignment, structural compliance, and a rigorous citation verification mechanism to prevent AI hallucinations. Experimental results indicate that agents trained within this closed-loop system demonstrate significant improvements in the depth of their reasoning and the accuracy of their citations. This approach seeks to overcome the inherent biases often found in subjective, model-based evaluation methods by providing a more reliable and verifiable assessment of scholarly work.