KMP-Bench (Dialogue) - Modelling: leaderboard
Metric: Modelling principle accuracy (%): win rate averaged over the three modelling-specific criteria on the turns that target this principle, on KMP-Dialogue: the tutor model continues truncated multi-turn K-8 math tutoring dialogues under a dialogue-specific system prompt naming the target pedagogical principle, and a Gemini-2.0-Flash judge compares each response with the original tutor turn (Win/Tie/Lose); higher is better. Source: arxiv.org. Saturation forecast: Estimated already saturated. 16 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2.5 72B Instruct | 71.5 | #436 |
| 2 | Claude 3.7 Sonnet | 71.2 | #241 |
| 3 | Qwen 2.5 32B Instruct | 70.7 | #491 |
| 4 | Phi-4 | 70 | #701 |
| 5 | Qwen 2.5 14B Instruct | 68.2 | #634 |
| 6 | Qwen 2.5 7B Instruct | 66.5 | #846 |
| 7 | GPT-4o | 57.5 | #333 |
| 8 | Mistral Nemo Instruct (2407) | 54.8 | #916 |
| 9 | Llama 3.3 70B Instruct | 48.8 | #520 |
| 10 | Qwen2.5-Math-72B-Instruct | 42.9 | #708 |
| 11 | Qwen2.5-Math-7B-Instruct | 34.5 | #1422 |
| 12 | Gemini 2.0 Flash | 34.2 | #331 |
| 13 | Llama 3.1 8B Instruct | 29 | #1018 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #8 Claude Fable 5.1.
Interactive version: theaggregate.ai/benchmark?slug=kmp-bench-dialogue-modelling · How It Works · Data refreshed daily, snapshot 2026-10-11.