AI RESEARCH
More Rounds, More Noise: Why Multi-Turn Review Fails to Improve Cross-Context Verification
arXiv CS.CL
•
ArXi:2603.16244v1 Announce Type: new Cross-Context Review (CCR) improves LLM verification by separating production and review into independent sessions. A natural extension is multi-turn review: letting the reviewer ask follow-up questions, receive author responses, and review again. We call this Dynamic Cross-Context Review (D-CCR). In a controlled experiment with 30 artifacts and 150 injected errors, we tested four D-CCR variants against the single-pass CCR baseline.