WatchAct - Procedural Reasoning: leaderboard

Metric: Plan success rate (%; the VLM's high-level Pick/Place/Open/Close plan, predicted from a real human video and an instruction, run through a symbolic executor and checked against the task's success predicate; instance-weighted mean of the Imitation, Reversal and Temporal Sort tasks, 725 instances). Source: arxiv.org. Saturation forecast: Around September 2027. 9 models tracked.

Top models

#ModelScore
1GPT-5.429.6
2Gemini 3.1 Pro (Preview)29.5
3GPT-5.127.9
4Gemini 3 Flash22.8
5Qwen 3 VL 235B A22B Instruct10.2
6Qwen 3 VL 235B A22B (Thinking)8
7Gemini 3.1 Flash Lite6.9

Interactive version: theaggregate.ai/benchmark?slug=watchact-procedural-reasoning · How It Works · Data refreshed daily, snapshot 2026-09-26.