Lunheng — Multi-Agent Paper Review (论衡) v2.0

"权衡论证, 校释虚妄" — 王充《论衡》(~80 CE)

v2.0 change log (2026-04-14): Added venue parameter + word-budget contract + D9 Narrative dimension + Chief Editor role + cuts field. v1.0 produced overstuffed papers (14500 chars, 16 \paragraph small-headings, fragmented FAQ-style narrative) because all 8 dimensions rewarded additions without a counter-force. v2.0 adds a budget-enforcing Chief Editor after Refiner and mandates every Evaluator propose cuts alongside fixes.

Operate on the paper at: $ARGUMENTS

What This Skill Does

Lunheng treats paper improvement as a contract-governed multi-agent process with six roles:

1. Architect (A_arch) — reads paper → emits Blueprint (incl. word-budget allocation) + Visual Contract
2. Evaluator Bench (E_1 ... E_9) — 9 parallel sub-agents, one per dimension; each emits fixes + cuts
3. Writer (A_w) — re-drafts sections under budget + contract constraints
4. Refiner (A_ref) — global polish + cross-ref + terminology harmonization
5. Chief Editor (A_ed) — [NEW in v2.0] enforces venue word-budget; can override evaluator additions to preserve narrative flow
6. Aggregator — composite score $R = \frac{1}{9} \sum D_k$ + budget-check, decides loop continuation

Constants

• MAX_ITERATIONS = 3
• QUALITY_THRESHOLD = 7.0 — composite score
• BUDGET_SLACK = 1.10 — allowed overshoot of venue word budget
• WORKSPACE = <paper_dir>/lunheng_workspace/
• CONTRACT_PATH = <paper_dir>/lunheng_workspace/visual_contract.json
• BUDGET_CONTRACT_PATH = <paper_dir>/lunheng_workspace/word_budget.json
• LOG = <paper_dir>/LUNHENG_LOG.md

Venue Profiles (required parameter)

venue	Main text (CN chars)	Abstract	Max \paragraph	Small-headings style
journal (default for JMC A / CEJ / Digital Discovery / JACS)	8,000 – 10,000	200-350	≤ 3	Connected paragraphs
nature_sub (Nat Comm / Nat Comp Sci / Nat Energy)	2,500 – 3,500	≤ 250 words (EN)	0	Flowing prose only
conference (NeurIPS / ICML / ICLR)	6,000 – 8,000 words (EN)	150-200 words	≤ 2	Connected paragraphs
thesis (CUPB / 本科 / 硕士毕业)	15,000 – 30,000	400-800	allowed	Sub-headings OK

STOP_CRITERION = (R ≥ QUALITY_THRESHOLD) AND (word_count ≤ venue_budget × BUDGET_SLACK)

If R passes but budget fails → skip Writer/Evaluator, run Chief Editor only to trim, then re-check.

The 9 Dimensions

ID	Dimension	Aligned with
D1	Soundness	NeurIPS Soundness, Nature Technical Rigor
D2	Significance & Originality	NeurIPS Significance + Originality
D3	Clarity & Organization	NeurIPS Clarity, JACS Clarity
D4	Experimental Substance	NeurIPS Quality (claims supported)
D5	Reproducibility	NeurIPS 16-item Checklist
D6	Citation Quality	JACS Literature Appropriateness
D7	Visual & Tabular Communication	Nature Figure Quality
D8	Ethics, Limitations & Broader Impact	NeurIPS Limitations + Ethics
D9	Narrative Flow & Conciseness [NEW v2.0]	venue-appropriate length + readable prose

D9 Rubric anchors (NEW)

• 10: Within 100% of venue budget; ≤ recommended \paragraph count; every cut would damage a claim
• 8-9: Within 110% of budget; ≤ 1.5× recommended small-headings; minor tightening possible
• 7: Within 125% of budget OR minor narrative fragmentation from 1-2 excess \paragraph blocks
• 5-6: Within 150% of budget OR reads as FAQ/spec sheet rather than prose
• 3-4: > 150% of budget OR section flow broken by >5 excess small-headings
• 1-2: ≥ 2× budget or unreadable as single-narrative paper

D3 Clarity — budget-aware amendments

Add these D3 anchors to the existing rubric:

• -2 on D3 if sub-section introduces redundancy already covered in prior section
• -1 on D3 if \paragraph{} title repeats content headline instead of structural transition

Workflow

Step 0: Initialize + Venue Budget Contract

mkdir -p <paper_dir>/lunheng_workspace
cp <paper_dir>/main.tex <paper_dir>/lunheng_workspace/main_round0.tex
cp <paper_dir>/main.pdf <paper_dir>/lunheng_workspace/main_round0.pdf 2>/dev/null || true

Parse venue from $ARGUMENTS. If missing, ASK user before proceeding. Do not default silently.

Write <paper_dir>/lunheng_workspace/word_budget.json:

{
  "venue": "journal",
  "main_budget_cn_chars": 9000,
  "main_budget_tolerance": 9900,
  "abstract_budget_cn_chars": 300,
  "max_paragraph_directives": 3,
  "per_section_budget": {
    "Abstract": 300,
    "Introduction": 1200,
    "Methods": 2200,
    "Results": 2500,
    "Discussion": 1200,
    "Conclusion": 400,
    "Data_Availability_and_Impact": 200
  }
}

Step 1: Architect Agent (prompt updated for v2.0)

You are the ARCHITECT agent for Lunheng v2.0.

Inputs:
- Paper: [paper_dir]/[main_tex]
- Figures: [paper_dir]/figures/
- Venue budget: <word_budget.json> (read first)

Produce THREE outputs:

1. blueprint.md — structured outline with per-section WORD BUDGET:
   - Ordered sections {s_1, s_2, ...} with argument outline
   - Key claims and supporting evidence
   - **Per-section target word count (must sum to main_budget)**
   - **Recommended use/non-use of \paragraph{} sub-headings per section**
   - 3 most severe structural issues

2. visual_contract.json — schema unchanged from v1.0.

3. word_budget.json — update per_section_budget if paper structure differs from template.

Report a 200-word summary:
- 3 structural issues
- Current vs budget word count delta per section
- Which sections will need compression vs expansion

Step 2: Round 1 Evaluator Bench (9 evaluators with CUTS)

Spawn all 9 evaluators in a single message. Each returns:

dimension: D_k
score: <1-10>
confidence: <1-5>
rubric_anchor_matched: "..."
strengths: [≥3]
weaknesses:
  CRITICAL: [...]
  MAJOR: [...]
  MINOR: [...]
fixes:   # additions
  - 动作: "..."
    位置: "..."
    估计增字: <N>   # [NEW v2.0] words this adds
    预期效果: "..."
cuts:    # [NEW v2.0] mandatory deletions to balance budget
  - 动作: "删除/压缩 X"
    位置: "..."
    估计省字: <N>
    依据: "为何这条对 claim 非必需"
score_change_criteria: [...]

D9 evaluator additionally returns:

word_budget_check:
  current: <N>
  target: <N>
  overshoot_pct: <%>
  paragraph_count: <N>
  max_allowed_paragraph: <N>

Aggregate $R_1 = \frac{1}{9} \sum D_k$. Save to LUNHENG_LOG.

Step 3: Generate-Evaluate-Adapt Loop

For iteration = 1 .. MAX_ITERATIONS:

If R ≥ QUALITY_THRESHOLD AND word_count ≤ budget × 1.10: STOP → Step 6. If R ≥ QUALITY_THRESHOLD BUT word_count > budget × 1.10: skip to Step 3d (Chief Editor only). Else: full Writer → Refiner → Chief Editor → Evaluator.

Step 3a: Writer agents (per-section, parallel) — v2.0 constraints

You are the WRITER agent for section "<s_i>".

Inputs:
- Blueprint constraint (incl. word budget for this section): <budget>
- Visual Contract
- Current section text
- Reviewer feedback (fixes + cuts for this section)

HARD CONSTRAINTS (v2.0):
1. Final section word count MUST be ≤ budget × 1.10
2. Each "fix" may add at most 100 words to the section. If the fix needs more,
   move the details to SI and cite "详见 SI §S<X>".
3. You MUST execute every "cut" proposed by evaluators unless cutting harms a claim.
4. Do NOT introduce new \paragraph{} sub-headings. The blueprint specifies whether \paragraph is allowed for this section; default is NO.
5. Prefer connected paragraphs over enumerate/itemize for lists ≤ 4 items.
6. Reference every visual in section_obligations; 1-2 sentences of context per \ref.
7. Use exactly the terms from terminology_glossary.
8. **[NEW v2.1] Numerical grounding (anti-placeholder)**: For every specific number, count, percentage, or ratio you emit, you MUST cite its source file path (e.g., `data/xxx.csv` row count, `output/yyy.json` field, computed from `zzz.npz`). If the source is inaccessible, either (a) use qualitative language ("majority", "small fraction") or (b) flag as `[VERIFY: <description>]` for Refiner/Chief Editor to resolve. **NEVER fabricate a specific count by multiplying total × rounded percentage** (e.g., `3555 × 60% = 2142` as a subset size) — doing so creates fake-specific numbers that imply experiments never actually run.

Produce:
- Revised LaTeX wrapped in <SECTION_OUTPUT>...</SECTION_OUTPUT>
- Word count reported: <before> → <after>
- Which cuts applied / which cuts refused with reason
- SI delta: list of items moved to SI

Step 3b: Refiner agent (unchanged from v1.0)

Merges all Writer outputs into draft_round<N>.tex, harmonizes terminology, verifies \ref / \cite integrity. Does not compress.

Step 3c: Chief Editor agent [NEW v2.0]

You are the CHIEF EDITOR for Lunheng v2.0.

Inputs:
- Refined draft: <refined_round<N>.tex>
- Word budget contract: <word_budget.json>
- Venue: <venue>
- Evaluator cuts compilation (from Step 2)

Your authority:
1. Override any Writer output that exceeded section budget
2. Delete \paragraph{} sub-headings added by Writer
3. Merge short paragraphs into flowing prose
4. Move method-detail, protocol, parameter-dump content to SI
5. Delete hedged/speculative passages that a venue reviewer would flag as padding
6. Condense Limitations enumerate to flowing prose (1 para, max 120 CN chars per limit)

Constraints:
- Preserve every core claim, every number, every \cite, every \ref
- Do NOT delete figures or tables; only in-text descriptions may be condensed
- Maintain terminology_glossary consistency

Output:
- editor_round<N>.tex (compressed draft)
- edit_log_round<N>.md (what was cut, by how much, where moved)
- Updated word_budget.json with actual per_section counts

Report:
- Before/after word count per section
- Total char reduction %
- Sections still over budget (for next iteration)

Step 3d: Evaluator re-run (fresh context, includes D9)

Step 4: Compile & Verify (unchanged)

Step 5: Final Visual Contract Validation (unchanged)

Step 6: Document Results + Budget Audit

LUNHENG_LOG.md now includes:

## Budget Compliance
- Venue: <journal/thesis/nature_sub/conference>
- Target main text: <N> CN chars
- Final main text: <N> CN chars
- Overshoot: <%>
- \paragraph count: <N> (max allowed: <M>)
- D9 Narrative score: <N>/10

Key Rules (v2.0 updates in bold)

• NEVER fabricate experimental results.
• PRESERVE all raw agent outputs in lunheng_workspace/.
• FAIL LOUD on contract violations.
• Sub-agent isolation.
• Parallel evaluators.
• Anchored scoring.
• [NEW v2.0] Budget is contract, not suggestion — Chief Editor enforces at every round.
• [NEW v2.0] Every fix must be paired with a cut of comparable size unless the paper is currently under budget. No net additions above budget slack.
• [NEW v2.0] \paragraph{} is opt-in per section — default is banned for journal/nature_sub/conference venues.
• [NEW v2.0] Writer cannot override Chief Editor — order of authority: Architect > Chief Editor > Refiner > Writer.
• [NEW v2.1] Writer must ground every number in a source file — no percentage-derived fake counts (e.g., total × rounded_pct → N in prose), no numbers pulled from memory without citing source. Any unverifiable number must be flagged [VERIFY: ...] for Refiner/Chief Editor resolution. Applies to counts, percentages, R²/ρ/MAE, EF, OOF metrics, etc.

Override Examples

• /lunheng paper/ venue:journal — standard journal compression
• /lunheng paper/ venue:thesis — thesis mode, budget relaxed
• /lunheng paper/ venue:journal max_iterations:2 — shorter loop
• /lunheng paper/ venue:journal stop_threshold:8.5 — higher quality bar
• /lunheng paper/ venue:nature_sub — Nature-style compression
• /lunheng paper/ venue:journal skip_editor:true — v1.0 compat mode (NOT RECOMMENDED)

When to Use vs. lunheng-quick

• lunheng (this skill): paper needs structural restructuring + length calibration. Costs ~12–30 sub-agent calls per round.
• lunheng-quick: paper is structurally OK, just needs polish. 2 rounds, ~4–6 sub-agent calls total. (Also inherits v2.0 budget awareness.)

Anti-patterns Lunheng v2.0 Prevents (learned from v1.0)

1. "Answer every reviewer question in-text" → v2.0 moves answers to SI
2. "Expand Limitations to 6 enumerate points to hit D8 10-band" → v2.0 D9 penalizes; Chief Editor flattens to prose
3. "Add \paragraph{协议} for every method detail" → v2.0 bans \paragraph by default
4. "Broader Impact 1000+ chars to defend against Ethics reviewer" → v2.0 hard-caps via budget
5. "Every fix adds 300 words, no fix ever subtracts" → v2.0 requires paired cuts
6. [NEW v2.1] "Percentage × total → fake-specific N in prose" (e.g., writing "the literature-only subset (n=2142)" when no such experiment was ever run, derived from a rounded 3555 × 60%) → v2.1 forces Writer to ground each number in a source file or tag [VERIFY: ...]. Motivation: Paper1 audit (2026-04-17) cleared 17 such placeholders — B subset n=2142, C subset n=3591, EF=1.57, occupancy 50.4%/32.1%, CC backfill 38%, SI CC equation form, N content unit, MAE 0.37, etc. — all from Writer's rounded-percentage arithmetic that never touched source data.