auto-paper-improvement-loop（自动纸张改进循环）

Auto Paper Improvement Loop: Review → Fix → Recompile

Autonomously improve the paper at: $ARGUMENTS

Context

This skill is designed to run after Workflow 3 (/paper-plan → /paper-figure → /paper-write → /paper-compile). It takes a compiled paper and iteratively improves it through external LLM review.

Unlike /auto-review-loop (which iterates on research — running experiments, collecting data, rewriting narrative), this skill iterates on paper writing quality — fixing theoretical inconsistencies, softening overclaims, adding missing content, and improving presentation.

Constants

• MAX_ROUNDS = 2 — Two rounds of review→fix→recompile. Empirically, Round 1 catches structural issues (4→6/10), Round 2 catches remaining presentation issues (6→7/10). Diminishing returns beyond 2 rounds for writing-only improvements.
• REVIEWER_MODEL = gpt-5.4 — Model used via Codex MCP for paper review.
• REVIEWER_BIAS_GUARD = true — When true, every review round uses a fresh mcp__codex__codex thread with no prior review context. Never use mcp__codex__codex-reply for review rounds. Set to false only for deliberate debugging of the legacy behavior. Empirical evidence (April 2026): running the same paper with codex-reply + "since last round we did X" prompts inflated scores from real 3/10 → fake 8/10 across 5 rounds; switching to fresh threads recovered the true 3/10 assessment.
• REVIEW_LOG = PAPER_IMPROVEMENT_LOG.md — Cumulative log of all rounds, stored in paper directory.
• HUMAN_CHECKPOINT = false — When true, pause after each round's review and present score + weaknesses to the user. The user can approve fixes, provide custom modification instructions, skip specific fixes, or stop early. When false (default), runs fully autonomously.

💡 Override: /auto-paper-improvement-loop "paper/" — human checkpoint: true

Inputs

1. Compiled paper — paper/main.pdf + LaTeX source files
2. All section .tex files — concatenated for review prompt

State Persistence (Compact Recovery)

If the context window fills up mid-loop, Claude Code auto-compacts. To recover, this skill writes PAPER_IMPROVEMENT_STATE.json after each round:

{
  "current_round": 1,
  "threadId": "019ce736-...",
  "last_score": 6,
  "status": "in_progress",
  "timestamp": "2026-03-13T21:00:00"
}

On startup: if PAPER_IMPROVEMENT_STATE.json exists with "status": "in_progress" AND timestamp is within 24 hours, read it + PAPER_IMPROVEMENT_LOG.md to recover context, then resume from the next round. Otherwise (file absent, "status": "completed", or older than 24 hours), start fresh.

After each round: overwrite the state file. On completion: set "status": "completed".

Reviewer Independence Protocol

The reviewer must be context-naive on every round. Prior-round summaries, fix lists, and executor explanations are not evidence; they are a source of confirmation bias. If the reviewer is told what changed, scores tend to drift upward even when the manuscript itself has not materially improved.

Rules:

• Every round starts with mcp__codex__codex, not mcp__codex__codex-reply.
• Never pass a prior threadId into the next review prompt.
• Never include "since last round", "we fixed", "after applying", or any fix summary in the reviewer prompt.
• The only acceptable evidence of improvement is the current .tex source and compiled PDF.
• If a fix cannot be observed in the files, the reviewer should not be told it happened.
• If recovery metadata is needed, store the returned threadId for crash recovery only; do not use it to preserve review context.

Set REVIEWER_BIAS_GUARD = false only if you explicitly want the legacy, context-carrying behavior for debugging.

Workflow

Step 0: Preserve Original

cp paper/main.pdf paper/main_round0_original.pdf

Step 1: Collect Paper Text

Concatenate all section files into a single text block for the review prompt:

# Collect all sections in order
for f in paper/sections/*.tex; do
    echo "% === $(basename $f) ==="
    cat "$f"
done > /tmp/paper_full_text.txt

Step 2: Round 1 Review

Send the full paper text AND compiled PDF to GPT-5.4 xhigh:

mcp__codex__codex:
  model: gpt-5.4
  config: {"model_reasoning_effort": "xhigh"}
  prompt: |
    You are reviewing a [VENUE] paper. Please provide a detailed, structured review.

    ## Paper Files:
    - LaTeX source: [list all section .tex files]
    - Compiled PDF: paper/main.pdf
    - Figures: [list figure files]

    Read BOTH the LaTeX source (for content/logic) AND the compiled PDF (for visual presentation).

    ## Review Instructions
    Please act as a senior ML reviewer ([VENUE] level). Provide:
    1. **Overall Score** (1-10, where 6 = weak accept, 7 = accept)
    2. **Summary** (2-3 sentences)
    3. **Strengths** (bullet list, ranked)
    4. **Weaknesses** (bullet list, ranked: CRITICAL > MAJOR > MINOR)
    5. **For each CRITICAL/MAJOR weakness**: A specific, actionable fix
    6. **Missing References** (if any)
    7. **Visual Review** (from the PDF):
       - Figure quality: readable? labels legible? colors distinguishable in grayscale?
       - Figure-caption alignment: does each caption match its figure?
       - Layout: orphaned headers, awkward page breaks, figures far from references?
       - Table formatting: aligned columns, consistent decimals, bold for best results?
       - Visual consistency: same color scheme across all figures?
    8. **Verdict**: Ready for submission? Yes / Almost / No

    Focus on: theoretical rigor, claims vs evidence alignment, writing clarity,
    self-containedness, notation consistency, AND visual presentation quality.

Save the threadId for Round 2.

Step 2b: Human Checkpoint (if enabled)

Skip if HUMAN_CHECKPOINT = false.

Present the review results and wait for user input:

📋 Round 1 review complete.

Score: X/10 — [verdict]
Key weaknesses (by severity):
1. [CRITICAL] ...
2. [MAJOR] ...
3. [MINOR] ...

Reply "go" to implement all fixes, give custom instructions, "skip 2" to skip specific fixes, or "stop" to end.

Parse user response same as /auto-review-loop: approve / custom instructions / skip / stop.

Step 3: Implement Round 1 Fixes

Parse the review and implement fixes by severity:

Priority order:

1. CRITICAL fixes (assumption mismatches, internal contradictions)
2. MAJOR fixes (overclaims, missing content, notation issues)
3. MINOR fixes (if time permits)

Common fix patterns:

Step 4: Recompile Round 1

cd paper && latexmk -C && latexmk -pdf -interaction=nonstopmode -halt-on-error main.tex
cp main.pdf main_round1.pdf

Verify: 0 undefined references, 0 undefined citations.

Step 4.5: Restatement Regression Test

After every recompilation, rerun a theorem-statement consistency check so fix rounds cannot reintroduce appendix drift. Run this after Step 4 and again after Step 7 before the final format check.

Scope

• Compare only theorem/lemma/proposition/corollary statements, not proof bodies.
• Classify files by main.tex input order: files before \appendix are main body; files after \appendix are appendix.

Normalized comparison logic

• Strip comments, \label{...}, \ref{...}, \eqref{...}, \cite...{...}, and whitespace-only differences.
• Collapse formatting-only macros such as \emph{}, \textbf{}, \textit{}, \mathrm{}, \mathbf{}, \mathcal{}, and \operatorname{} to their contents.
• Preserve quantifiers, case splits, assumptions, and the literal names of defined objects.
• Compare by theorem label when available; otherwise compare by theorem type and order.
• Flag any change in hypotheses, case splits, quantifier order, or terminology (stationary vs terminal) as regression drift.

python3 - <<'PY'
import re
def normalize(s):
    s = re.sub(r'%.*', '', s)
    s = re.sub(r'\\label\{[^}]*\}', '', s)
    s = re.sub(r'\\(?:ref|eqref|cref|Cref|cite[a-zA-Z]*)\{[^}]*\}', '', s)
    s = re.sub(r'\\(?:emph|textbf|textit|mathrm|mathbf|mathsf|mathcal|operatorname)\{([^{}]*)\}', r'\1', s)
    s = re.sub(r'\\begin\{[^}]+\}|\\end\{[^}]+\}', '', s)
    s = re.sub(r'\s+', ' ', s)
    return s.strip().lower()
# Compare normalized theorem blocks from the current main-body files
# against their appendix restatements. Any mismatch blocks completion.
PY

Empirical motivation: in our April 2026 NeurIPS run, thm:dsm-oracle had a 3-case split (w=0/1/>1) in main but no case split in appendix; nu_T was named "stationary" in main and "terminal" in appendix. These drifted multiple times across fix rounds because no automated check caught regression.

Step 5: Round 2 Review

If REVIEWER_BIAS_GUARD = true (default), use a fresh mcp__codex__codex thread for Round 2. Do not reuse the Round 1 threadId for prompting. Save the returned threadId only for recovery bookkeeping.

mcp__codex__codex:
  model: gpt-5.4
  config: {"model_reasoning_effort": "xhigh"}
  prompt: |
    You are reviewing a [VENUE] paper. This is a fresh, zero-context review.
    Ignore any prior review rounds, prior fix lists, or executor explanations.
    Judge the paper only from the current LaTeX source and compiled PDF.

    ## Paper Files:
    - LaTeX source: [list all section .tex files]
    - Compiled PDF: paper/main.pdf
    - Figures: [list figure files]

    Read BOTH the LaTeX source (for content/logic) AND the compiled PDF (for visual presentation).

    ## Review Instructions
    Please act as a senior ML reviewer ([VENUE] level). Provide:
    1. **Overall Score** (1-10, where 6 = weak accept, 7 = accept)
    2. **Summary** (2-3 sentences)
    3. **Strengths** (bullet list, ranked)
    4. **Weaknesses** (bullet list, ranked: CRITICAL > MAJOR > MINOR)
    5. **For each CRITICAL/MAJOR weakness**: A specific, actionable fix
    6. **Missing References** (if any)
    7. **Visual Review** (from the PDF):
       - Figure quality: readable? labels legible? colors distinguishable in grayscale?
       - Figure-caption alignment: does each caption match its figure?
       - Layout: orphaned headers, awkward page breaks, figures far from references?
       - Table formatting: aligned columns, consistent decimals, bold for best results?
       - Visual consistency: same color scheme across all figures?
    8. **Verdict**: Ready for submission? Yes / Almost / No

    Focus on: theoretical rigor, claims vs evidence alignment, writing clarity,
    self-containedness, notation consistency, and visual presentation quality.

If REVIEWER_BIAS_GUARD = false (legacy debugging only), use mcp__codex__codex-reply with the saved threadId; this is not the recommended path.

Step 5.5: Kill Argument Exercise (theory papers only)

Run this only if the paper is theory-heavy (≥5 \begin{theorem}|\begin{lemma}|\begin{proposition}|\begin{corollary} environments in the source) and only on the final scheduled round (current_round == MAX_ROUNDS).

This is a late-stage adversarial check. It must always use fresh mcp__codex__codex threads, never codex-reply, and it must not reuse any prior review context.

Thread 1: Attack

• Use a fresh thread with only the current paper files.
• Prompt: "Construct the single best argument to reject this paper in 200 words. Focus on theorem validity, assumption mismatch, missing proof obligations, limit-order ambiguity, and claim/evidence gaps. Do not reference prior rounds or fixes."

Thread 2: Defense

• Use a second fresh thread with the current paper files plus the attack memo.
• Prompt: "Now defend the paper against the attack memo. For each rejection point, classify it as already fixed, partially fixed, or still unresolved, and cite the current files. Do not reuse prior review context."

Merge rule

• Dedupe attack points against the Round 2 weakness list by semantic overlap.
• Append any novel unresolved attack point to the Step 6 fix list before implementation.
• If the defense cannot refute a point, keep it at the original severity or raise it by one level if it exposes a main-theorem or core-assumption failure.
• If the defense shows the issue is already fixed in the current files, only downgrade after verifying the file evidence.
• Record both memos in PAPER_IMPROVEMENT_LOG.md.
• If HUMAN_CHECKPOINT = true, include the merged findings in the checkpoint summary before asking the user to proceed.

This phase feeds directly into Step 6. The attack/defense findings must be merged before the final recompile.

Empirical motivation: in our April 2026 NeurIPS run, after 5 rounds of standard improvement (score 7-8/10), the kill-argument exercise surfaced framing weaknesses that no prior review caught (e.g., "width-w is mostly conditional", "CRF irrelevant to real D-LLMs"). Author rebuttal forced explicit scope qualifications in abstract and discussion.

Step 5b: Human Checkpoint (if enabled)

Skip if HUMAN_CHECKPOINT = false. Same as Step 2b — present Round 2 review, wait for user input.

Step 6: Implement Round 2 Fixes

Same process as Step 3. Typical Round 2 fixes:

• Add controlled synthetic experiments validating theory
• Further soften any remaining overclaims
• Formalize informal arguments (e.g., truncation → formal proposition)
• Strengthen limitations section

Step 7: Recompile Round 2

cd paper && latexmk -C && latexmk -pdf -interaction=nonstopmode -halt-on-error main.tex
cp main.pdf main_round2.pdf

Step 8: Format Check

After the final recompilation, run a location-aware format compliance check.

# If the log lacks file/line data, rerun the final compile once with -file-line-error.
cd paper && latexmk -pdf -file-line-error -interaction=nonstopmode -halt-on-error main.tex

# 1. Page count vs venue limit
PAGES=$(pdfinfo paper/main.pdf | grep Pages | awk '{print $2}')
echo "Pages: $PAGES (limit: 9 main body for ICLR/NeurIPS)"

# 2. Duplicate labels: HARD BLOCK
DUP_LABELS=$(grep -Rho "\\\\label{[^}]*}" paper/main.tex paper/sections 2>/dev/null | sort | uniq -d || true)
if [ -n "$DUP_LABELS" ]; then
    echo "Duplicate labels found (BLOCKING):"
    echo "$DUP_LABELS"
fi

# 3. Overfull warnings with location classification
OVERFULLS=$(grep -n "Overfull \\\\hbox" paper/main.log 2>/dev/null || true)

# Main body = source files before \appendix in main.tex.
# Appendix = source files after \appendix, or files whose path contains "appendix".
# Bibliography = paper.bbl, references.bib, or bibliography-generated output.
MAIN_BODY_OVERFULL=$(echo "$OVERFULLS" | grep -v -E 'appendix|paper\.bbl|references\.bib' || true)
APPENDIX_OVERFULL=$(echo "$OVERFULLS" | grep -E 'appendix' || true)
BIB_OVERFULL=$(echo "$OVERFULLS" | grep -E 'paper\.bbl|references\.bib' || true)

echo "Main-body overfulls (any size BLOCKS):"
echo "$MAIN_BODY_OVERFULL"
echo "Appendix overfulls (>10pt blocks):"
echo "$APPENDIX_OVERFULL"
echo "Bibliography overfulls (>20pt blocks):"
echo "$BIB_OVERFULL"

Stop criteria:

• Any duplicate label blocks completion.
• Any overfull in the main body blocks completion, regardless of size.
• Appendix overfulls block completion only if they exceed 10pt or are visibly clipping.
• Bibliography overfulls block completion only if they exceed 20pt or are visibly clipping.
• Underfull hboxes remain warnings unless they create obvious layout damage.

Auto-fix patterns (location-aware):

Location-aware interpretation:

• Classify by the source file reported in the -file-line-error log.
• If a warning cannot be classified, treat it as main body and fix it.

Empirical motivation: in our April 2026 NeurIPS run, 28+ overfull hbox warnings (largest 160pt in the appendix bridge proof) survived 5 improvement rounds because the previous blanket "overfull > 10pt blocks" rule was too lax and treated all locations equally.

Step 9: Document Results

Create PAPER_IMPROVEMENT_LOG.md in the paper directory:

# Paper Improvement Log

## Score Progression

| Round | Score | Verdict | Key Changes |
|-------|-------|---------|-------------|
| Round 0 (original) | X/10 | No/Almost/Yes | Baseline |
| Round 1 | Y/10 | No/Almost/Yes | [summary of fixes] |
| Round 2 | Z/10 | No/Almost/Yes | [summary of fixes] |

## Round 1 Review & Fixes

<details>
<summary>GPT-5.4 xhigh Review (Round 1)</summary>

[Full raw review text, verbatim]

</details>

### Fixes Implemented
1. [Fix description]
2. [Fix description]
...

## Round 2 Review & Fixes

<details>
<summary>GPT-5.4 xhigh Review (Round 2)</summary>

[Full raw review text, verbatim]

</details>

### Fixes Implemented
1. [Fix description]
2. [Fix description]
...

## PDFs
- `main_round0_original.pdf` — Original generated paper
- `main_round1.pdf` — After Round 1 fixes
- `main_round2.pdf` — Final version after Round 2 fixes

Step 9: Summary

Report to user:

• Score progression table
• Number of CRITICAL/MAJOR/MINOR issues fixed per round
• Final page count
• Remaining issues (if any)

Feishu Notification (if configured)

After each round's review AND at final completion, check ~/.claude/feishu.json:

• After each round: Send review_scored — "Round N: X/10 — [key changes]"
• After final round: Send pipeline_done — score progression table + final page count
• If config absent or mode "off": skip entirely (no-op)

Output

paper/
├── main_round0_original.pdf    # Original
├── main_round1.pdf             # After Round 1
├── main_round2.pdf             # After Round 2 (final)
├── main.pdf                    # = main_round2.pdf
└── PAPER_IMPROVEMENT_LOG.md    # Full review log with scores

Key Rules

• Large file handling: If the Write tool fails due to file size, immediately retry using Bash (cat << 'EOF' > file) to write in chunks. Do NOT ask the user for permission — just do it silently.
• Preserve all PDF versions — user needs to compare progression
• Save FULL raw review text — do not summarize or truncate GPT-5.4 responses
• Reviewer independence (Round 2+): when REVIEWER_BIAS_GUARD = true (default), use a fresh mcp__codex__codex thread for every review round; never use mcp__codex__codex-reply and never include "since last round" / fix summaries in the prompt. See the Reviewer Independence Protocol section above.
• Always recompile after fixes — verify 0 errors before proceeding
• Do not fabricate experimental results — synthetic validation must describe methodology, not invent numbers
• Respect the paper's claims — soften overclaims rather than adding unsupported new claims
• Global consistency — when renaming notation or softening claims, check ALL files (abstract, intro, method, experiments, theory sections, conclusion, tables, figure captions)

Typical Score Progression

Based on end-to-end testing on a 9-page ICLR 2026 theory paper:

+4.5 points across 3 rounds (2 content + 1 format) is typical for a well-structured but rough first draft. Final: 8 pages main body, 0 overfull hbox, ICLR-compliant.

Review Tracing

After each mcp__codex__codex or mcp__codex__codex-reply reviewer call, save the trace following shared-references/review-tracing.md. Use tools/save_trace.sh or write files directly to .aris/traces/<skill>/<date>_run<NN>/. Respect the --- trace: parameter (default: full).