DiscoBench - Guided - Accuracy: leaderboard
Correct final answers in the interactive search task. Interactive deep search with ambiguity checkpoints, Tavily search and Gemini 3 Flash Medium user simulation. Table 2 of paper v2; Neutral and Guided prompts stay separate. Model effort settings follow the API appendix. GPT-5.4 has no Guided score. This series reports End-to-end accuracy (%). Source: https://arxiv.org/html/2606.27669v2.
Metric: End-to-end accuracy (%). Source: arxiv.org. Saturation forecast: Around August 2027. 10 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) (High) | 53.1 | #64 (Gemini 3.1 Pro (Preview)) |
| 2 | Doubao-Seed-2-0-Pro-260215 (High) | 50.2 | |
| 3 | DeepSeek V4 Pro (Max) | 38.9 | #112 (DeepSeek V4 Pro) |
| 4 | Claude Opus 4.7 (Max) | 38.9 | #51 (Claude Opus 4.7) |
| 5 | GLM-5.1 (Thinking) | 38.4 | #109 (GLM-5.1) |
| 6 | Kimi K2.6 (Thinking) | 35.1 | #113 (Kimi K2.6) |
| 7 | MiMo-V2.5-Pro (Thinking) | 28.4 | #167 (MiMo-V2.5-Pro) |
| 8 | Hy3-preview (High) | 24.2 | #166 (Hy3-preview) |
| 9 | Qwen 3.6 Max (Preview) (Thinking) | 14.6 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=discobench-guided-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-05.