Goal-Conditioned Supervised Learning for LLM Fine-Tuning

In three linesNew offline fine-tuning method for LLMs: Goal-Conditioned Supervised Learning (GCSL) treats feedback signals as explicit goals and trains models via pure supervised learning. Evaluated on non-toxic generation, code generation, and recommendation; outperforms SFT and DPO without external reward models.

Summary generated by Claude — human-verified