WatchAct - Implicit Intent Inference: 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 Nonverbal Cue and Reference Disambiguation tasks, 455 instances). Source: arxiv.org. Saturation forecast: Around July 2027. 9 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview)32.4
2Gemini 3 Flash27.8
3GPT-5.420.7
4Qwen 3 VL 235B A22B (Thinking)12.6
5GPT-5.112.4
6Gemini 3.1 Flash Lite6.4
7Qwen 3 VL 235B A22B Instruct5.3

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