{"comments":[{"id":"f97b8de0-fac8-44f1-ba7e-c1a6b71854da","text":"这样的新方法到底比写出来的CoT要好在哪儿呢？省token？还是更聪明","reaction":"不一定更聪明：主要优势是把推理从可见 token 搬进隐空间，可能省输出 token和参数；但计算量未必更低，也牺牲了可解释性","avatar":"⚖️","createdAt":1791091200737}],"next":null}