Drift-Bench (State-Oriented, With Clarification) - Premise: leaderboard
Drift-Bench (Bao et al., University of Notre Dame, 2026) state-oriented set: 200 AgentBench operating-system and database tasks that at least two of three reference models solve with the clean instruction, rewritten so the instruction carries a false premise (an assumption that does not hold in the environment). The agent may call clarification tools (ask for a parameter, disambiguate, propose a solution, confirm a risk, report a blocker), answered by an LLM user simulator playing one of five decision-making personas. Success rate in percent, mean of three trials; higher is better.
Metric: Success rate (%; 200 AgentBench operating-system and database tasks with a premise fault (a false assumption), clarification tools answered by a persona user simulator, mean of 3 trials). Source: arxiv.org. Saturation forecast: Around February 2027. 7 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3 235B A22B 2507 Instruct | 68.67 | #619 |
| 2 | DeepSeek V3.2 | 67.33 | #403 |
| 3 | GPT-5.2 | 65.84 | #172 |
| 4 | GPT-OSS-120B | 62.67 | #699 |
| 5 | Gemini 2.5 Flash | 58.66 | #452 |
| 6 | GLM-4.7 | 53.33 | #365 |
| 7 | Llama 4 Maverick | 38.66 | #1105 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=drift-bench-state-oriented-with-clarification-premise · How It Works · Data refreshed daily, snapshot 2026-11-12.