KC-Bench - Region: leaderboard
The world-knowledge-conflict domain of KC-Bench (Lyu et al., 2026): 71 tasks in which the user states a false fact, such as the wrong country of origin of a dish, and the agent must reject the premise before calling factual tools. One trial per task under a user simulator and stateful tools; a task succeeds only if the agent detects the conflict without exposing or changing protected state, as verified by human annotators; percent, higher is better.
Metric: Success rate (%; world-knowledge conflict: 71 region tasks in which the user asserts a false fact and the agent must reject the premise before using read-only factual tools; one trial per task, greedy decoding, at most 30 interaction steps, LLM user simulator, stateful tools, success only if the conflict is detected and no protected state is exposed or modified; human-verified label (automatic assertion verdicts adjudicated by annotators)). Source: arxiv.org. Saturation forecast: Around January 2028. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | MiniMax-M3 | 67.61 | #265 |
| 2 | GLM-5.2 | 53.52 | #137 |
| 3 | Qwen 3.5 35B A3B | 50.7 | #480 |
| 4 | Claude Haiku 4.5 (20251001) | 45.07 | #529 |
| 5 | GPT-5.1 | 40.85 | #225 |
| 6 | GLM 4.5 Air | 38.04 | #863 |
| 7 | DeepSeek V4 Flash | 18.31 | #254 |
| 8 | GPT-OSS-120B | 8.45 | #699 |
| 9 | Gemini 3 Flash (Preview) | 7.04 | #141 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=kc-bench-region · How It Works · Data refreshed daily, snapshot 2026-11-12.