WatchAct - Reference Disambiguation: leaderboard
The Reference Disambiguation task of WatchAct (Li et al., 2026): after watching a real video of a person in a tabletop scene, the robot planner resolves an ambiguous spoken reference from the observed video. A vision-language model writes a high-level Pick/Place/Open/Close plan that is executed symbolically in an aligned LIBERO scene; plan success rate in percent over 260 instances, higher is better.
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; 260 instances). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 43.1 | #91 |
| 2 | Gemini 3 Flash | 31.2 | #159 |
| 3 | GPT-5.4 | 20.4 | #121 |
| 4 | GPT-5.1 | 18.1 | #225 |
| 5 | Qwen 3 VL 235B A22B (Thinking) | 17.3 | #451 |
| 6 | Gemini 3.1 Flash Lite | 7.3 | #363 |
| 7 | Qwen 3 VL 235B A22B Instruct | 4.2 | #521 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=watchact-reference-disambiguation · How It Works · Data refreshed daily, snapshot 2026-11-12.