Seems interesting to let the LLMs design their own reasoning traces instead of being constrained by human labelers. I could imagine some self consistency approaches to find common high-quality reasoning traces.
Seems like a bitter lesson moment for reasoning traces.
1 comment
[ 3.9 ms ] story [ 15.8 ms ] threadSeems like a bitter lesson moment for reasoning traces.
PDF of the Paper - https://arxiv.org/pdf/2509.06160