benchmarks evals
The Best Robot Explored Too Little
RoboQuest's strongest multimodal agent completed only 23% of tasks requiring search, inspection or interactive testing.
Summary
RoboQuest's strongest multimodal agent completed only 23% of tasks requiring search, inspection or interactive testing.
The benchmark separates hidden-information discovery from basic execution. Agents could often perform the component actions when the missing fact was supplied, yet commonly committed before gathering required evidence and rarely repaired disturbances caused during exploration. A policy fine-tuned on full episodes almost never succeeded.
Why it matters
RoboQuest's strongest multimodal agent completed only 23% of tasks requiring search, inspection or interactive testing.
Limits and context
No additional limitation was separately recorded.
Key claims
RoboQuest's strongest multimodal agent completed only 23% of tasks requiring search, inspection or interactive testing.
Evidence: source-2026-10-08-018
Sources
- arXiv preprint 2610.10388arXiv · primary research
Corrections
No corrections have been recorded for this story.