What problem does it solve? First-draft academic papers often contain theoretical inconsistencies, overclaims, notation confusion, and formatting violations that require multiple rounds of expert review to fix. This Skill automates the review-fix-recompile cycle so a generated paper is polished by an external LLM reviewer without manual coordination. ## Core Features & Use Cases - Two-Round Review Loop: Sends the full concatenated LaTeX source to GPT-5.4 (xhigh reasoning) via Codex MCP, receives a structured score and ranked weaknesses, implements fixes by severity, and recompiles with latexmk. - State Persistence & Recovery: Writes PAPER_IMPROVEMENT_STATE.json after each round so the loop can resume after context compaction, and preserves every PDF version (round 0, 1, 2) for comparison. - Format Compliance Check: Detects overfull/underfull hbox warnings and page-limit violations, then applies auto-fix patterns such as \resizebox and content moves to the appendix. - Use Case: After generating an ICLR-style paper with the paper-write and paper-compile workflow, run this Skill to raise the review score from roughly 4/10 to 7/10 across two autonomous rounds while logging every change in PAPER_IMPROVEMENT_LOG.md. ## Quick Start Run the auto paper improvement loop on the compiled paper in the paper/ directory to review, fix, and recompile it for two rounds.
