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.
More from Research
- Tobias Lee: RL for Verifiable Tasks, MOPD for Open Domains — _AndrewZhao · 2026-09-17
- Schmidhuber Publishes 40-Year Retrospective on Recursive Self-Improvement Since 1987 — rbhar90 · 2026-09-17
- LocalFold runs AlphaFold3, Boltz-2 and more protein folding models in your browser via WebGPU — sokrypton · 2026-09-17
- ColabFold2 preview unifies AlphaFold3, Boltz2, RosettaFold3 and more in JAX — sokrypton · 2026-09-17
- Doctor: AI is crushing maths but has barely touched clinical trials and medicine — LaCaipirinha · 2026-09-17
- ColabFold 1.6.3 Released: 2.5x Faster, pip-Installable, Adds ipSAE+pDockQ2 Scores — sokrypton · 2026-09-17