SFT Mitigates Mode Collapse and Over-Dispersion in LLMs

Supervised fine-tuning (SFT) can drive LLM output diversity toward the target distribution, potentially fixing mode collapse. The study derives a bias-variance decomposition to show that the degree of mode collapse depends on specific model and dataset combinations.