What problem does it solve? After generating a research paper draft, authors face the tedious cycle of getting critical feedback, applying fixes, and recompiling — often missing theoretical inconsistencies, overclaims, and formatting issues that reviewers would catch. This Skill automates that polish loop with an external LLM reviewer. ## Core Features & Use Cases - Two-round review-fix-recompile loop: Sends LaTeX source and compiled PDF to GPT-5.5 (xhigh reasoning) for a structured senior-reviewer critique, implements fixes by severity (CRITICAL > MAJOR > MINOR), and recompiles with latexmk. - Reviewer independence guard: Every round uses a fresh review thread with no prior context, preventing score inflation from confirmation bias. - Edit whitelist constraints: Optionally restrict fixes to specific paths and forbid operations like new citations, theorem environments, or numerical claims — essential for resubmit and camera-ready modes. - Quality gates: Includes theorem restatement regression checks, duplicate-label detection, overfull-hbox format checks, and an optional adversarial kill-argument exercise for theory-heavy papers. - Use Case: After compiling a NeurIPS submission draft, run the loop to autonomously raise the paper from a 4/10 to a 7/10 review score across two rounds, with all changes logged in PAPER_IMPROVEMENT_LOG.md. ## Quick Start Run the auto-paper-improvement-loop on my compiled paper in the paper/ directory to review, fix, and recompile it for two rounds.
