MMCL-Bench - Rule System Application: leaderboard

MMCL-Bench (2026) tests multimodal context learning: a model must learn a task-local rule, procedure or pattern from diagrams, screenshots, manuals or videos and apply it to a new visual instance, graded against expert rubrics. This board is the strict pass rate on the 49 rule-system tasks, where a task counts only if every rubric item holds; percent, higher is better.

Metric: Strict task pass rate (%) on the 49 rule-system tasks (a task passes only when every rubric item is satisfied), on the 102 MMCL-Bench tasks (49 rule-system, 32 procedural, 21 empirical), in which a model learns a task-local rule, procedure or pattern from visual or mixed-modality teaching context and applies it to a new visual instance, scored against expert rubrics by a rubric-only LLM judge, models at their recommended or default decoding settings; higher is better. Source: arxiv.org. Saturation forecast: Around 2028. 5 models tracked. Updated 7 Oct 2026.

Top models

#ModelSettingScoreOverall rank
1GPT-5.4Thinking32.7#135
2Qwen 3.6 PlusThinking28.6#238
3Kimi K2.5Thinking20.4#266
4Claude Opus 4.6Thinking20.4#113
5Gemini 3.1 Pro (Preview)16.3#99

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

Interactive version: theaggregate.ai/benchmark?slug=mmcl-bench-rule-system-application · How It Works · Data refreshed daily, snapshot 2026-11-21.