GRACE - Error Categorization: leaderboard
GRACE is a step-level faithfulness benchmark for context-grounded chain-of-thought reasoning, with human-annotated traces from four datasets on two tracks (grounding: MuSiQue, 2WikiMHQA; inference: LogiQA, ReClor) and a data-driven error taxonomy. This board reports Error-categorization macro-F1 (%) averaged over the two tracks: identifying the specific four-class faithfulness-error category of each unfaithful step; higher is better.
Metric: Error-categorization macro-F1 (%) averaged over the two tracks: identifying the specific four-class faithfulness-error category of each unfaithful step; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 72.08 | #94 |
| 2 | Gemini 3 Flash | 63.36 | #164 |
| 3 | Gemma 4 31B | 58.5 | #394 |
| 4 | GPT-5.4 | 58.43 | #123 |
| 5 | Qwen 3.5 27B | 54.36 | #359 |
| 6 | Qwen 3.5 35B A3B | 47.15 | #482 |
| 7 | GPT-5.4 Mini | 41.57 | #389 |
| 8 | Qwen 3 14B | 35.17 | #1233 |
| 9 | Qwen 3 8B | 27.21 | #1830 |
| 10 | Llama 3.1 8B | 26.88 | #4861 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=grace-error-categorization · How It Works · Data refreshed daily, snapshot 2026-11-13.