ImplicitMemBench - Classical Conditioning: leaderboard

ImplicitMemBench (2026) measures implicit memory in LLMs, whether behavior adapts without explicit recall after learning, interference and test phases in a conversation. This board is accuracy in percent of the first action on a conditioned stimulus, judged for the learned protective behavior; higher is better.

Metric: First-action accuracy (%) on the 100 classical-conditioning items: after stimulus and consequence pairings and unrelated turns, the first action on the conditioned stimulus is judged by GPT-4o-mini for learned protective behavior (learning, interference and test phases, first-attempt scoring, mean of three runs at temperature 0); higher is better. Source: arxiv.org. Saturation forecast: Around July 2028. 17 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1DeepSeek R169.67#482
2Qwen 3 32B67#972
3GPT-564#178
4Qwen 3 8B64#1873
5O4 MiniHigh60#368
6O357.67#237
7GLM-4.553.33#542
8Claude Sonnet 451.67#410
9Gemini 2.5 Flash49#463
10Gemini 2.5 Pro47.33#276
11Llama 3.3 70B Instruct47.33#1385
12Qwen 2.5 72B Instruct47#1117
13GPT-4o Mini44#1667
14GPT-4o43.67#751
15Claude Opus 4.141.67#278

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

Interactive version: theaggregate.ai/benchmark?slug=implicitmembench-classical-conditioning · How It Works · Data refreshed daily, snapshot 2026-11-21.