Compatibility-Aware Dynamic Fine-Tuning for Large Language Models

In three linesCADFT (Compatibility-Aware Dynamic Fine-Tuning) extends Dynamic Fine-Tuning by controlling sample-level variance. The method uses a compatibility signal derived from model likelihoods to modulate supervised updates and suppresses high-variance gradients from incompatible demonstrations. Experiments show improved stability, generalization, and RL cold-start initialization.

Summary generated by Claude — human-verified