Sep-Latent: Separator Tokens for Optimizing Implicit Reasoning in LLMs
In brief
- Sep-Latent uses separator tokens as semantic anchors for implicit reasoning.
- The method cuts verbosity to 5.4 tokens on average while achieving 56.8% accuracy on GSM8k-Aug.
- A multi-level distillation mechanism ensures logical consistency between explicit and implicit reasoning.
A new method, Sep-Latent, enhances language model efficiency by compressing implicit reasoning into separator tokens. It reduces verbosity while maintaining accuracy close to explicit approaches like Chain-of-Thought.
This summary gives you the essentials. The full article, its figures and quotes are on cs.AI updates on arXiv.org.
Read the original articleAI-generated summary of the source, checked by rules. Report an error