Most times I read an RL chain of thought it feels more human to me than the output text.
Like, I go “Wow, that feels more closely like a reflection of my own thought than the output or non RLVR’d output. The leaps feel human-thought-leap sized; they don’t feel as polished and like they are making an argument published for public consumption. It’s like when I have been caught in some particular train of thought for a while, and started making little shortcuts inside.”
I am reporting a feeling, not proposing a conclusion about what this feeling means.
Most times I read an RL chain of thought it feels more human to me than the output text.
Like, I go “Wow, that feels more closely like a reflection of my own thought than the output or non RLVR’d output. The leaps feel human-thought-leap sized; they don’t feel as polished and like they are making an argument published for public consumption. It’s like when I have been caught in some particular train of thought for a while, and started making little shortcuts inside.”
I am reporting a feeling, not proposing a conclusion about what this feeling means.
Are you talking about the raw CoT? I haven’t gotten this feeling (much) from the summarized CoTs I’ve seen in Claude.