模型感知 token 预算后思维链更压缩
AI 导读
我觉得这实际上可能是改进后的 CoT 可控性在真实场景中的一个实例——模型越注意到自己"预算快用完了"，思考就变得越压缩 @tomekkorbak @MicahCarroll
正文
I think this is actually might be an instance of improved CoT controllability in the wild -- the more model notices that it is "runs out of budget" more compressed the thinking becomes
@tomekkorbak @MicahCarroll
Our original attack allows extracting reasoning of the recent frontier models, including Astra and Sol 6.1.
On reasoning effort MAX both Astra and Sol become very aware of their "token budgets", and eventually start saving tokens by omitting white spaces
More examples on https://stolen-thoughts.com/在 X 查看被引用的帖子