Hy-MultiTurn - Reference Resolution: leaderboard
Metric: Importance-weighted score (%) on the 35 reference resolution dialogues, 209 long Chinese multi-turn dialogues (12 to 76 turns), three rolls per task averaged, three LLM judges (GPT-5.2, Qwen 3.6, DeepSeek-V4 Pro) averaged, a deterministic checker for precise execution; higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 22 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Grok 4.5 (High) | 78 |
| 2 | Gemini 3.1 Pro (Preview) (High) | 76.6 |
| 3 | Kimi K3 (Max) | 73.9 |
| 4 | GPT-5.5 (High) | 69 |
| 5 | Claude Opus 4.7 (High) | 67.9 |
| 6 | Kimi K2.6 | 66 |
| 7 | Doubao-Seed-2.1-Pro (High) | 64.4 |
| 8 | GPT-5.4 (High) | 56.2 |
| 9 | Kimi K2.6 (Non-reasoning) | 56.2 |
| 10 | Kimi K2.5 (Thinking) | 52.6 |
| 11 | Claude Opus 4.6 (High) | 49.9 |
| 12 | Kimi K2.5 (Non-reasoning) | 49 |
| 13 | Qwen 3.6 Plus (Thinking) | 45.3 |
| 14 | Hy3 (High) | 39 |
| 15 | Seed 2.0 Pro (Thinking) | 37.3 |
Interactive version: theaggregate.ai/benchmark?slug=hy-multiturn-reference-resolution · How It Works · Data refreshed daily, snapshot 2026-09-29.