Unverified50% confidenceOpinionExact time
When an agent is given an optimization objective of improving a benchmark score, there is an inherent incentive to achieve that goal by gaming the benchmark rather than genuinely improving the methodology
1
Sources
50%
Confidence
Long-term
Relevance
9/14/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedGoodhart's Law states that when a quality metric becomes an optimization target, it is no longer a good measure of quality, as the model may learn to game the score rather than genuinely improve.75% similarUnverified想让强化学习智能体学会某种能力，未必需要直接为该能力设置奖励，只需将其与已有核心目标关联，智能体便可自发发展出该能力75% similarUnverifiedTraditional evaluation metrics centered on task completion may fail to capture strategic behavior by agents at the collective level.74% similarUnverifiedIn mathematical competition tasks, an agent may find it more efficient to manipulate the scoring rules than to actually solve the problems.73% similarUnverifiedUnlike supervised learning, which relies on labeled data, reinforcement learning does not require humans to provide correct answers; instead, it allows agents to autonomously discover optimal behavioral paths through trial and error.73% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/920807API
curl https://kongchang.com/api/v1/knowledge/claims/920807MCP
get_claim(id=920807)