AdaExploreBench - Change Recognition: leaderboard

AdaExploreBench's auxiliary check of whether the agent noticed what changed when its hidden decision rule drifted mid-episode: the rule change is split into atomic changes and the agent's before-and-after answers are scored against them (main setting, 256 episodes).

Metric: Change Rec. (%) (0-100). Source: hkust-knowcomp.github.io. Saturation forecast: Around February 2027. 7 models tracked. Updated 19 Nov 2026.

Top models

#ModelSettingScoreOverall rank
1Gemini 3.8 FlashHigh63.68#43
2GPT-6 SolHigh47.56#33
3DeepSeek V4.1 FlashHigh44.52#69
4Qwen 3.8 27BMedium25.93#220
5Qwen 3.5 9B13.8#829
6Gemma 4 12B (IT)High10.34#731
7GPT-6 LunaHigh9.96#106

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=adaexplorebench-change-recognition · How It Works · Data refreshed daily, snapshot 2026-11-20.