When2Tool (Multi-Hop): leaderboard
Multi-hop tool-use decisions from When2Tool (Sun et al., 2026): 450 test tasks in three simulated environments, each a chain of three dependent tool calls, at easy, medium and hard difficulty. Accuracy in percent of exact-form answers checked by a deterministic evaluator, with tools offered under a neutral default prompt, mean of three seeds; higher is better.
Metric: Accuracy (%; exact-form answers in a boxed final response, checked by a deterministic evaluator; simulated offline tool environments at easy, medium and hard difficulty; 450 multi-hop test tasks in 3 environments, each a chain of three dependent tool calls; tools offered under the neutral default prompt; mean of three seeds). Source: arxiv.org. Saturation forecast: Estimated already saturated. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 32B | 88.9 | #960 |
| 2 | Qwen 3 14B | 87.5 | #1235 |
| 3 | Qwen 3 4B 2507 Instruct | 82.1 | #2174 |
| 4 | Llama 3.3 70B Instruct | 62.4 | #1353 |
| 5 | Llama 3.1 8B | 40.2 | #4900 |
| 6 | Qwen 3 1.7B | 21.2 | #5551 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=when2tool-multi-hop · How It Works · Data refreshed daily, snapshot 2026-11-12.