UserToolBench - Single-Tool: leaderboard
Single-tool tasks of UserToolBench (Sun Yat-sen University and Huawei Noah Ark Lab, 2026): the model must pick the right tool and fill the decision-relevant arguments that the hidden user profile implies, inferring them from the interaction history. Exact trajectory accuracy in percent; higher is better.
Metric: Exact trajectory accuracy (%; single-tool tasks; a trajectory counts when it matches the profile-conditioned reference: the tool name and decision-relevant arguments, the ordered call sequence for multi-tool tasks, or the correct clarify-or-infer decision when information is missing; 799 profile-hidden evaluation instances from long-horizon interaction trajectories of 10 privacy-sanitized user profiles over 36 tool sets; the model sees the interaction history, the current request and the tool schemas but not the profile). Source: arxiv.org. Saturation forecast: Not forecast. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GLM-5 | 70.62 | #260 |
| 2 | DeepSeek V4 Pro | 61.95 | #188 |
| 3 | Qwen 3.6 Plus | 58.02 | #222 |
| 4 | Kimi K2.6 | 52.3 | #178 |
| 5 | Gemini 3.5 Flash | 52.24 | #101 |
| 6 | GPT-5.4 | 40.22 | #123 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=usertoolbench-single-tool · How It Works · Data refreshed daily, snapshot 2026-11-13.