Establishing Objective Reward Functions for Peer Review
Traditional peer review suffers from high variance, subjectivity, and a lack of standardized reward signals for AI agents tasked with evaluation. The InternReviewer and InternAdvocate framework addresses this by introducing a reinforcement learning (RL) architecture that treats the review process as a multi-agent game. By decoupling the 'reviewer' agent from the 'advocate' (rebuttal) agent, the system creates a competitive environment that forces both parties to ground their arguments in the actual content of the paper rather than relying on hallucinated critiques or generic praise.
The Agentic Feedback Loop
The framework utilizes a dual-agent structure to refine evaluation quality:
- InternReviewer: Tasked with identifying substantive flaws and providing constructive, evidence-based feedback. Its reward function is calibrated to prioritize accuracy and alignment with the paper's claims.
- InternAdvocate: Acts as the author-proxy, generating rebuttals to the reviewer's critiques. This forces the reviewer agent to defend its claims, effectively creating a self-correcting loop where weak or biased reviews are penalized if they cannot withstand logical scrutiny.
By iterating through this rebuttal process, the system generates a more robust 'objective reward' signal. This signal is used to train the agents, moving away from static prompt-based evaluation toward a dynamic, policy-based approach that improves over time through simulated interaction.