Fine-tuning fixes LLM mode collapse and over-dispersion, new arXiv paper shows

chrmanning · x · 2026-09-17

An arXiv paper formalizes LLM output diversity via sequence collision probability and shows mode collapse is not inevitable: whether it occurs depends on model, dataset, and post-training. With enough SFT data, diversity converges to the target distribution, the gap is bounded by the square root of KL divergence, and fine-tuning can fix both under- and over-diversity.

Original post →

More from Research

Research channel →