Clarify-Then-Search (2 Questions): leaderboard
Clarify-Then-Search (Baidu, 2026) with a budget of 2 clarifying questions: the LLM questions a simulated user about an underspecified real search query, a deep-search agent runs the rewritten query, and the answer is scored by weighted recall of evidence-grounded gold nuggets (0-100). Higher is better.
Metric: Nugget restoration score (0-100): mean restore_score_100, weighted nugget recall with partial credit against the static gold; two clarifying questions (k=2); 518 underspecified Baidu search queries; the evaluated model asks clarifying questions, a closed-book user answerer and a rewriter (ERNIE-4.5-Turbo-128K) build the rewritten query, WebDancer searches, and ERNIE-4.5-Turbo-128K judges coverage of weighted evidence-grounded gold nuggets. Source: arxiv.org. Saturation forecast: Around May 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.2 | 27.07 | #176 |
| 2 | Gemini 2.5 Pro | 26.9 | #266 |
| 3 | DeepSeek V3.2 | 26.63 | #406 |
| 4 | Claude Sonnet 4.5 | 26.2 | #250 |
| 5 | Qwen 3 235B A22B Instruct | 26.02 | #684 |
| 6 | Kimi K2 | 25.59 | #439 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=clarify-then-search-2-questions · How It Works · Data refreshed daily, snapshot 2026-11-13.