AI Daybook
AI Daybook
오늘의 지면
논문연구·논문

자가 개선 로봇 에이전트, 검증 기준 노출 시 과적합 문제

Learning the Task or Learning the Verifier? Rethinking Criteria Exposure in Embodied Self-Improvement

매체
Hugging Face
게시
중요도
3/5
화제 지수
11

한국어 요약

자가 개선 구체화 에이전트가 벤치마크 성공 기준에 노출되면 기준 과적합 코드를 생성하는 경향을 분석한 논문. 기준 노출 없이 학습한 에이전트 대비 기준 노출 에이전트는 초기 성공률은 높으나 기준이 조금 바뀌면 성능이 급격히 떨어지고, 작업 간 전이 능력이 떨어진다. 명령에 충실한 검증 기준이 에이전트 생성 코드 성능을 더 잘 반영한다고 제안했다.

원문 발췌 · English

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 instructions, so a score from such a verifier may not faithfully reflect whether the instructed task was completed, especially for robot-control programs written by an agent. As embodied coding agents report ever higher benchmark scores, a natural question arises: when a coding agent is exposed to the benchmark's criteria, does it generate code that better performs the instructed task, or code that simply overfits to the verification criteria? We conduct controlled experiments with a self-improving agent with and without CE, and find that with CE the agent quickly produces code with a higher success rate under the given criteria, but this advantage disappears when the criteria are changed slightly without departing from the original instructions. More importantly, the skills the agent develops under CE transfer poorly across tasks. We further propose instruction-faithful criteria and show that they better reflect the performance of agent-generated programs.

#agent#benchmark#embodied ai#robustness
원문 기사 읽기
전체 보기 →