Learning the Task or Learning the Verifier? Rethinking Criteria Exposure in Embodied Self-Improvement
Learning the Task or Learning the Verifier? Rethinking Criteria Exposure in Embodied Self-Improvement
This research paper examines the impact of 'Criteria Exposure' (CE) on self-improving embodied agents. The authors find that while exposing agents to benchmark verification criteria allows them to quickly generate code with higher success rates under those specific criteria, this advantage vanishes when the criteria are slightly altered. Furthermore, skills learned under CE transfer poorly across different tasks. The study proposes 'instruction-faithful criteria' as a better metric for reflecting the actual performance of agent-generated programs.
The study investigates whether exposing embodied agents to benchmark verification criteria leads to genuine task performance or overfitting, finding that criterion-exposed agents achieve higher scores on specific tests but fail to transfer skills when criteria change slightly.
Source: arXiv Robotics — research abstracts · Read original article ↗
Loading article text…
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S2
- paper: not_reported
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
arXiv:2611.08318v1 Announce Type: cross Abstract: Self-improving embodied agents are increasingly being developed and compared on robot simulation benchmarks. Several recent works expose the benchmark's success criteria to the learning agent, a practice we call Criteria Exposure (CE). These criteria, originally written to verify policies trained by imitation learning, often demand more than the corresponding language
Open source S2
Source：arXiv Robotics — research abstracts · arxiv.org