AI-powered manuscript review and revision system inspired by APRES (ICLR 2026). Evaluates scientific manuscripts using ScholarEval 8-dimension rubric plus citation-predictive heuristics, then performs targeted revisions while preserving core scientific claims. Outputs before/after comparison with improvement metrics. Use when the user says "/review", "帮我审一下", "review my manuscript", "improve this paper", "polish this draft", or provides a manuscript for quality improvement. Also triggered by "审稿", "修改论文", "润色".
Evaluate and improve scientific manuscripts through a closed-loop process: Score → Identify Weaknesses → Revise → Re-score → Compare. Inspired by APRES (Meta Superintelligence Labs, ICLR 2026).
/review or /review <path-to-manuscript>Score the manuscript on each dimension (0.00–1.00):
| Dimension | Weight | Evaluation Criteria |
|---|---|---|
| Novelty | 15% | Does this advance knowledge? Are claims clearly differentiated from prior work? |
| Rigor | 25% | Methodology sound? Statistics correct? Controls adequate? Sample sizes reported? |
| Clarity | 10% | Writing clear? Figures self-explanatory? Logical flow between sections? |
| Reproducibility | 15% | Methods detailed enough to replicate? Software versions stated? Data accessible? |
| Impact | 20% | Does this matter for the field? Broad or narrow implications? |
| Coherence | 10% | Do all parts fit together? Introduction → Methods → Results → Discussion aligned? |
| Limitations | 3% | Are limitations honestly acknowledged? Not buried or trivialized? |
| Ethics | 2% | Ethical standards met? IRB mentioned if applicable? Conflicts disclosed? |
Compute weighted average. Output initial verdict: accept (≥0.75), minor_revision (≥0.60), major_revision (≥0.40), reject (<0.40).
Evaluate 10 presentation factors that predict higher citation impact:
| # | Factor | Check | Weight |
|---|---|---|---|
| 1 | Title specificity | Contains key finding or quantitative result? Not vague? | High |
| 2 | Abstract conclusion | Has a clear, quantitative take-home message? | High |
| 3 | Figure self-sufficiency | Can each figure be understood from its caption alone? | High |
| 4 | Methods reproducibility | Software versions, parameters, thresholds all stated? | Medium |
| 5 | Statistical reporting | Effect sizes + CIs alongside p-values? Test assumptions verified? | Medium |
| 6 | Discussion balance | Presents counter-arguments and alternative interpretations? | Medium |
| 7 | Limitations honesty | Dedicated section with specific (not generic) limitations? | Medium |
| 8 | Introduction funnel | Narrows from broad context → gap → specific question? | Low |
| 9 | Reference recency | Includes papers from last 2 years? Not relying on outdated reviews? | Low |
| 10 | Data availability | States where data/code can be accessed? | Low |
Score each 0–1. Flag factors scoring below 0.5 as revision targets.
Sort all identified weaknesses by estimated impact on manuscript quality. Output a numbered revision plan:
## 修订计划（按影响力排序）
1. [HIGH] 标题过于笼统 → 改为包含主要发现的具体标题
当前: "The Role of THBS2 in Cancer"
建议: "THBS2 Overexpression Associates with M2 Macrophage Infiltration and Poor Survival Across 17 Cancer Types"
理由: 具体标题平均被引用量高 22%（Paiva et al., 2012）
2. [HIGH] 摘要缺少定量结论 → 添加关键数字
当前: "THBS2 was significantly upregulated in multiple cancers"
建议: "THBS2 was significantly upregulated in 17/33 TCGA cancer types (Wilcoxon p<0.001), with highest expression in PAAD (HR=2.31, 95%CI: 1.45-3.68)"
理由: 摘要中包含具体数字的论文被引用量高 29%
3. [MEDIUM] Figure 2 caption 缺少统计方法说明
...
Apply each revision to the manuscript text. For each change:
Constraints:
Re-run ScholarEval on the revised manuscript. Output comparison:
## 审修效果对比
| Dimension | Before | After | Change |
|---------------|--------|-------|--------|
| Novelty | 0.72 | 0.72 | — |
| Rigor | 0.68 | 0.75 | +0.07 |
| Clarity | 0.55 | 0.78 | +0.23 |
| Reproducibility| 0.60 | 0.82 | +0.22 |
| Impact | 0.70 | 0.70 | — |
| Coherence | 0.65 | 0.80 | +0.15 |
| Limitations | 0.40 | 0.75 | +0.35 |
| Ethics | 0.90 | 0.90 | — |
| | | | |
| **Weighted** | **0.66**| **0.76** | **+0.10** |
| **Verdict** | minor_revision | accept | ⬆ |
Citation-impact heuristics: 4/10 → 8/10 factors above threshold
共执行 12 处修订，主要改善了清晰度（+0.23）和可复现性（+0.22）。
核心科学主张和数据未做任何修改。
Save revised manuscript and revision report: