PaperAudit-Bench (Fast Mode) - Grok 4 Errors: leaderboard
PaperAudit-Bench (Tu et al., 2026) tests whether a model finds errors injected into accepted papers. This board uses the 46 NeurIPS 2025 oral papers as corrupted by Grok 4 (10 to 20 injected errors of eight types each), reviewed by each model in the Fast mode's single pass over the whole paper; text-only models see the papers without figures. Macro-F1 in percent over papers, the harmonic mean of error coverage and finding precision with a GPT-5.1 judge matching findings to injected errors, higher is better; the synthesis model is itself one of the detectors.
Metric: Macro-F1 (%; the per-paper harmonic mean of error coverage and finding precision, averaged over the 46 NeurIPS 2025 oral papers of the benchmark's NeurIPS branch as corrupted by grok-4 with 10-20 injected errors of eight types; Fast mode, one global review pass over the whole paper; findings matched to the injected errors by a GPT-5.1 judge; the three text-only detectors read the papers with figures removed). Source: arxiv.org. Saturation forecast: Around March 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 2.5 Pro | 34.9 | #266 |
| 2 | Claude Sonnet 4.5 | 33.9 | #250 |
| 3 | GPT-5 | 33.8 | #167 |
| 4 | GLM-4.6 (Non-reasoning) | 29.6 | #486 (GLM-4.6) |
| 5 | Qwen 3 235B A22B 2507 Instruct | 29.4 | #619 |
| 6 | Grok 4 | 28.6 | #307 |
| 7 | O4 Mini | 27.7 | #353 |
| 8 | Kimi K2 0905 | 20.9 | #558 |
| 9 | DeepSeek V3.1 (Non-reasoning) | 19.2 | #497 (DeepSeek V3.1) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=paperaudit-bench-fast-mode-grok-4-errors · How It Works · Data refreshed daily, snapshot 2026-11-13.