← Back to feed

I asked LLMs which game 40 neural nets play, given only the weights

AI · Studies · 2026-11-04 16:24

作者复现“从权重猜训练游戏”实验，用12款游戏共36个新网络测试GPT-6 Astra与GPT-6.1 Sol，发现部分游戏能稳定骗过模型，另一些则容易识别。

Source: wohoef.github.io

Auto-generated summary for reference only. RSS