AILLM mode collapse and fine-tuning’s role in output diversity
A paper’s author says language models can produce too little or too much variety relative to a target distribution—and claims fine-tuning can correct both.
TLDR
A paper’s author says mode collapse—too little output diversity—is not inevitable: it depends on the dataset and post-training method. Models can sometimes be more diverse than the target distribution, the author says, and their paper claims fine-tuning fixes both under-diversity and over-diversity. The author describes measuring diversity through “sequence collision”: how often two responses to the same prompt match. When exact matches are too rare, they describe using higher-level similarity instead.
Combined views
83
1 Source, first seen 1d ago
LLM mode collapse and fine-tuning’s role in output diversity
A paper’s author says language models can produce too little or too much variety relative to a target distribution—and claims fine-tuning can correct both.