Found in Conversation: LLMs Teach Themselves to Close the Multi-Turn Gap

In three linesFound in Conversation (FiC) is a training framework where LLMs self-teach to close the multi-turn gap (Lost-in-Conversation). Via View-Asymmetric Self-Distillation, the model distills between single-turn (teacher) and multi-turn (student) views. Tested on Llama, Qwen, Phi, OLMo (3B-14B), FiC recovers 92-100% of single-turn performance.

Summary generated by Claude — human-verified