auto-review-loop-minimax

Automate iterative paper reviews by running an external MiniMax review loop that repeatedly collects feedback, applies action-item fixes, and re-reviews ML research until it is ready for submission. It handles prompt and response cycles autonomously up to a configurable maximum.

State persistence during context compaction is handled by saving the current progress to a review-stage/REVIEW_STATE.json file. The loop retrieves this state to resume multi-round reviews seamlessly, while a cumulative log is written to review-stage/AUTO_REVIEW.md.

You need MiniMax API access to run the autonomous research improvement loop. The system uses an MCP MiniMax chat tool when available, but it includes a resilient fallback to direct curl calls to the MiniMax chat completions endpoint if the MCP tool is absent.

The stopping criteria for an automated multi-round review loop are based on reviewer scores and verdicts. The loop stops early when the external reviewer indicates the ML research work is ready or almost ready for submission, preventing unnecessary review cycles.

You can track verbatim reviewer feedback across multiple rounds because the loop writes a cumulative, round-by-round log to review-stage/AUTO_REVIEW.md. This audit trail captures raw reviewer responses and tracks the applied action-item fixes.

Avoid using an autonomous review-to-fix-to-re-review loop when manual iteration is preferred or when you lack MiniMax API access. It is specifically designed for rigorous conference-style feedback on ML/AI paper development workflows requiring structured action-item implementation.