Fine-tuning may correct both too little and too much LLM output diversity

A paper's author says mode collapse is not inevitable: depending on the dataset and post-training method, LLMs can produce too little variation or sometimes exceed the target distribution's diversity. The author reports that fine-tuning fixes both under-diversity and over-diversity.

The author describes measuring diversity through “sequence collision” probability—how often two responses to the same prompt match. Because exact matches can be extremely rare, broader similarity measures can be useful. One experiment compared code using abstract syntax trees.