GRACE: leaderboard
Many reasoning tasks require models to reason over input context, from document-grounded question answering to rule-based deduction. Chain-of-Thought (CoT) prompting produces traces that appear transparent, yet individual steps can silently deviate from the source evidence, even when the final answer is correct. Existing methods detect hallucinations at the response level but fail to identify wher
Metric: Avg Step (self-reported). Source: benchmarklist.com. Saturation forecast: Around June 2027. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 81.46 | #58 |
| 2 | GPT-5.4 | 78.53 | #79 |
| 3 | Gemma 4 31B | 77.46 | #200 |
| 4 | Qwen 3.5 35B A3B | 75.7 | #256 |
| 5 | Gemini 3 Flash | 75.19 | #99 |
| 6 | Qwen 3.5 27B | 74.59 | #187 |
| 7 | GPT-5.4 Mini | 66.85 | #212 |
| 8 | Qwen 3 14B | 56.6 | #537 |
| 9 | Qwen 3 8B | 49.56 | #674 |
| 10 | Llama 3.1 8B | 45.76 | #1151 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=grace · How It Works · Data refreshed daily, snapshot 2026-10-20.