iPrefBench Interaction Precision
Agents · 2026-10-26
iPrefBench's interaction precision: the share of the agent's individual information requests, pooled across the 130 tasks, that elicit an ask-only requirement not already elicited; requests for inferable or irrelevant details and repeated requests count against it. The denominator is each agent's own request count, so it is a diagnostic of unnecessary asking rather than a fixed-item score, in percent.
Top models (higher is better)
| Model | Score |
|---|---|
| Gemini 3.1 Pro Preview | 77.1 |
| Kimi K3 | 75.0 |
| Gemini 3.7 Flash | 70.6 |
| Qwen3.8 27B | 70.6 |
| Claude Opus 5 | 65.8 |
| Sonnet 5 | 60.4 |
| Qwen3.8-Flash-Next | 60.4 |