The authors train a language model along explicit and implicit reasoning pathways.
Separator tokens compress hidden reasoning while adapting its depth
Sep-Latent concentrates reasoning information into punctuation-like anchors, distills explicit chains into latent states, and learns when to stop internal computation.
Chinese Tech
Yinlong Xu · Jianfeng Huang · Zhengdong Zhou · Mingjie Lin · Fuyu Lv · Dan Ou · +2 more
Zhejiang University · Alibaba Group
Research Digest··3 min read
Xu and colleagues propose Sep-Latent, a framework for replacing lengthy written chains of thought with a short, variable-length sequence of continuous hidden states.
§