TRACE Bench - Diversity: leaderboard
TRACE Bench response diversity: a normalized repetition penalty from character-bigram Jaccard similarity between each reply's sentences and the model's earlier sentences (TRACE Bench, arXiv 2608.11236v1, Table 8). Percent; higher means less repetition. An auxiliary response-level dimension of the overall score.
Metric: Normalized repetition-penalty score (%). Source: arxiv.org. Saturation forecast: Estimated already saturated. 26 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V4 Pro (Non-reasoning) | 95.43 | #112 (DeepSeek V4 Pro) |
| 2 | Gemma 4 31B (Non-reasoning) | 93.49 | #212 (Gemma 4 31B) |
| 3 | GPT-5.5 | 92.58 | #33 |
| 4 | Claude Sonnet 4.6 | 92.22 | #94 |
| 5 | DeepSeek V4 Flash | 91.79 | #155 |
| 6 | Ministral 3 14B | 90.11 | #664 |
| 7 | Qwen 3.5 27B (Non-reasoning) | 89.48 | #192 (Qwen 3.5 27B) |
| 8 | Kimi K2.5 (Non-reasoning) | 88.7 | #152 (Kimi K2.5) |
| 9 | GLM-5 (Non-reasoning) | 88.24 | #149 (GLM-5) |
| 10 | MiniMax-M2.7 | 87.82 | #270 |
| 11 | MiMo-V2.5-Pro | 87.59 | #167 |
| 12 | MiMo-V2.5 | 86.62 | #196 |
| 13 | Qwen 3.5 2B (Non-reasoning) | 78.61 | #1114 (Qwen 3.5 2B) |
| 14 | Qwen 3 14B (Non-reasoning) | 74.76 | #542 (Qwen 3 14B) |
| 15 | Qwen 3 8B (Non-reasoning) | 66.45 | #678 (Qwen 3 8B) |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=trace-bench-diversity · How It Works · Data refreshed daily, snapshot 2026-11-05.