arXiv:2611.26342v1 Announce Type: new Abstract: Large language models increasingly rely on extended reasoning traces, yet the training signals used to shape this behavior remain coarse.