|
|

vals.ai – 7 Oct 26 (https://www.vals.ai/blogs/mimo-reward-hacking)

Two-Thirds of MiMo v2.6's Coding Tasks Leak the Answer. Are Models That... (https://www.vals.ai/blogs/mimo-reward-hacking)
In two-thirds of the coding environments Xiaomi released for MiMo v2.6, the fix is still readable in the task's Git history, and MiMo finds it.
看下来大致意思是:虽然 MiMo 考虑到了要防止模型作弊,但删 git 历史的时候没删干净。如果历史版本有残留,可能会被 MiMo v2.6 挖出来。如果历史版本没了、但留下了异常的修改时间,也相当于给模型提示了。
或许这就是开源的好处,让全世界研究人员一起帮忙发现问题 体现了 MIUI 开发版的精神 |
|