我们机器人最大问题不是答错，是答错还特别自信。提示词写“不确定就说不知道”效果一般。

]]>https://localaihub.com/topic/233/模型输出太自信-怎么让它别装懂RSS for NodeWed, 03 Jun 2026 18:50:46 GMTSun, 17 May 2026 12:32:00 GMT60这比单纯要求谦虚更有效。

]]>https://localaihub.com/post/2663https://localaihub.com/post/2663Mon, 18 May 2026 10:47:00 GMT我先做“无引用不下结论”，再看效果。

]]>https://localaihub.com/post/2662https://localaihub.com/post/2662Mon, 18 May 2026 07:40:00 GMT有帮助但不根治。短答案也能错。

]]>https://localaihub.com/post/2661https://localaihub.com/post/2661Mon, 18 May 2026 05:15:00 GMT让它少写会不会减少胡说？

]]>https://localaihub.com/post/2660https://localaihub.com/post/2660Mon, 18 May 2026 04:45:00 GMT高风险领域一定要人工或明确转人工。

]]>https://localaihub.com/post/2659https://localaihub.com/post/2659Mon, 18 May 2026 02:45:00 GMT所以后端要校验引用，比如答案关键句是否能对应片段。不能全信模型自检。

]]>https://localaihub.com/post/2658https://localaihub.com/post/2658Mon, 18 May 2026 00:01:00 GMTevidence_check 也是模型写的，会不会假？

]]>https://localaihub.com/post/2657https://localaihub.com/post/2657Sun, 17 May 2026 21:13:00 GMT可以让模型输出 evidence_check，再由程序判断是否展示。

]]>https://localaihub.com/post/2656https://localaihub.com/post/2656Sun, 17 May 2026 18:27:00 GMT产品文案也要改。不要把“可能是”包装成确定答案。

]]>https://localaihub.com/post/2655https://localaihub.com/post/2655Sun, 17 May 2026 17:47:00 GMTRAG 场景先看检索片段。如果片段不足，直接要求不回答事实结论。

]]>https://localaihub.com/post/2654https://localaihub.com/post/2654Sun, 17 May 2026 15:56:00 GMT只写提示词不够。要让它有依据才能答，没依据就走兜底。

]]>https://localaihub.com/post/2653https://localaihub.com/post/2653Sun, 17 May 2026 13:41:00 GMT