최고 AI 에이전트도 DiscoBench에서 실패: 질문하기가 검색보다 나은 이유
tencent의 새로운 discobench는 gemini 3.1 pro와 claude opus 4.7 같은 최고 AI 모델조차 후속 질문을 하지 못한다는 사실을 드러냈으며, 이는 신뢰할 수 있는 에이전트에게 치명적인 공백입니다.
tencent's hunyuan team and tsinghua university introduced discobench to expose a critical weakness in ai search agents: they fail to ask clarifying questions when queries are ambiguous. the benchmark tests 11 models across 211 tasks with 463 ambiguous points, finding that top performers like gemini 3.1 pro and claude opus 4.7 still score below 50% accuracy. the research shows that while agents are good at searching syntax, they lack the 'uncertainty recognition' to pause and ask, causing error cascades in multi-hop reasoning. however, when models successfully ask targeted follow-up questions, success rates jump to over 93%. this signals that future ai systems must prioritize proactive dialogue over autonomous action to handle real-world complexity.