ClarifyCodeBench - Key Question Rate (TKQR): leaderboard
ClarifyCodeBench (Peking University, 2026) deletes one to three details from 419 LiveCodeBench v6 problems and lets the model ask clarification questions, answered by the benchmark when a GPT-4o judge matches them to the annotated key questions, before it writes code. This board is the turn-discounted key question rate, an nDCG-style 0 to 1 credit for asking the annotated key questions early; higher is better.
Metric: Turn-discounted key question rate (0 to 1; nDCG-style credit for asking the annotated key clarification questions early, normalized by the ideal order, on 419 ambiguous LiveCodeBench v6 tasks with one to three deleted requirement details; questions matched to the annotated ones by a GPT-4o judge, majority of three; at most 6 rounds, temperature 0). Source: arxiv.org. Saturation forecast: Around June 2028. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | DeepSeek V3.2 (Non-reasoning) | 0.3 | #406 (DeepSeek V3.2) |
| 2 | Gemini 2.5 Flash (Non-reasoning) | 0.29 | #453 (Gemini 2.5 Flash) |
| 3 | GPT-4o | 0.22 | #741 |
| 4 | GPT-5 | 0.21 | #167 |
| 5 | Claude Sonnet 4.5 | 0.12 | #250 |
| 6 | Claude Sonnet 4.5 (Thinking) | 0.12 | #250 (Claude Sonnet 4.5) |
| 7 | Qwen 3 235B A22B (Non-reasoning) | 0.07 | #623 (Qwen 3 235B A22B) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=clarifycodebench-key-question-rate-tkqr · How It Works · Data refreshed daily, snapshot 2026-11-13.