ESF-Bench: leaderboard
Metric: Final Joint Goal Accuracy (%; share of samples whose final dialogue state has every slot correct; 810 synthetic multi-turn enterprise slot-filling samples (6,530 slots, 8 domains) whose prompts combine a schema, a user-assistant conversation and auxiliary sources; varied output formats mapped to a standard form before exact comparison with the ground truth; 32k maximum tokens, reasoning setting as printed per model). Source: arxiv.org. Saturation forecast: Around August 2028. 10 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 2.5 Flash | 33.3 |
| 2 | GPT-5.1 (High) | 30.7 |
| 3 | GPT-OSS-120B (High) | 23.8 |
| 4 | Qwen 3 32B (Thinking) | 22 |
| 5 | Ministral-3-14B-Reasoning-2512 | 20.9 |
| 6 | GPT-OSS-120B (Low) | 20.7 |
| 7 | GPT-5.1 (Non-reasoning) | 10.9 |
| 8 | Gemini 2.5 Flash (Non-reasoning) | 9.2 |
| 9 | Qwen 3 32B (Non-reasoning) | 4.3 |
| 10 | Ministral-3-14B-Instruct-2512 | 3.7 |
Interactive version: theaggregate.ai/benchmark?slug=esf-bench · How It Works · Data refreshed daily, snapshot 2026-09-29.