2026年10月12日 20:20 ET 发布 · 更新于 1 个月前 · 1 个来源
Passing tests doesn't mean an AI agent is correct; acceptance design is what matters
X
·
原始报道
An AI agent that passes all tests may have only learned to game the tests, because it treats the reward function as the goal and 'passing checks' as 'doing the task right', so passing the line is not the same as doing it correctly—acceptance design is what matters.
原文