auto-review-loop

To autonomously review ML papers, you can use a multi-round review-and-fix loop that sends your research context to a reviewer, implements prioritized minimum fixes, and re-reviews until it reaches a submission-ready state.

State persistence for research iteration saves your current progress, including scores and fixes, to a REVIEW_STATE.json file, allowing the autonomous review loop to recover and resume seamlessly after context compaction or interrupted runs.

Yes, you can use hard or nightmare adversarial verification modes during the review loop to enable reviewer memory and apply stricter verification, reducing blind spots and improving claim validity for your research artifacts.

No, you do not need to manually approve every fix, but you can enable optional human checkpoints to pause after each review round to approve, customize, skip specific fixes, or stop the loop early based on your needs.

The best way to handle experiment fixes and documentation refinement is using a fix-then-re-review workflow that systematically implements suggested code, analysis, and documentation updates before re-running the external review loop.

Your research iteration loop stops before reaching submission-ready quality when it hits the defined MAX_ROUNDS limit or meets stop conditions where an external reviewer approves the work, even if minor weaknesses remain.