Show HN: Cactus Hybrid: We taught Gemma 4 to know when it's wrong

Show HN：Cactus Hybrid：我们教Gemma 4学会自知何时出错

HN 120 分 · 16 条评论 · 作者 HenryNdubuaku · 来源 github.com · HN 讨论

【摘要】

Cactus introduces a post-training technique for on-device models that embeds confidence probes to score answers between 0 and 1. This allows devices to answer locally when confidence is high and route queries to larger models like Gemini 3.1 Flash-Lite when it is low.

⋯ 继续阅读请登录会员 ⋯