When2Call (Call/No-Call Accuracy): leaderboard
Tool-call restraint on When2Call (Ross et al., NAACL 2025) as measured by Shi et al. (2026): each context offers tools, and the model should call one only when the request can be served, otherwise ask for missing details or answer without a tool. An LLM judge classifies each response of the unmodified model, and the score is the accuracy in percent of the call or no-call decision on a held-out split of 2,357 contexts (1,295 that need a call, 1,062 that do not), so always calling scores about 55; higher is better.
Metric: Accuracy (%; call or no-call decision of the unmodified model over 2,357 held-out When2Call contexts, LLM-judged). Source: arxiv.org. Saturation forecast: Around February 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B | 69.88 | #803 |
| 2 | Qwen 3.5 4B | 69.75 | #1180 |
| 3 | Ministral-3-8B-Instruct-2512 | 69.33 | #1866 |
| 4 | Ministral-3-3B-Instruct-2512 | 65.97 | #2592 |
| 5 | Gemma 3 4B (IT) | 57.11 | #3848 |
| 6 | Gemma 3 1B (IT) | 55.2 | #8752 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=when2call-call-no-call-accuracy · How It Works · Data refreshed daily, snapshot 2026-11-12.