Get Started
Home
Topics
Search
Library
Research questionHow can we measure whether generative models preserve conditional diversity without multiple reference outputs per input?A generative model may produce plausible outputs while covering fewer of the alternatives represented in its training data. Measuring this conditional diversity is difficult when each input has only one observed reference output.
AI
Evaluation & Benchmarks
Machine Learning
Natural Language Processing
Research Paper
Statistical Machine Learning
Latest papersRecent research connected to this question, newest first.Do Large Language Models Capture the Diversity in their Training Data?The evidence covers language models with publicly available training data, including OLMo, Pythia, and GPT-Neo, as well as class-conditioned ImageNet generators and text-conditioned MS-COCO models. It uses conditional entropy and a matrix-based von Neumann entropy analogue on paired input-output data, examining model scale, sequence length, and decoding strategy; the proposed correction reweights multiple generated outputs through an entropy-constrained projection.research paper · Sep 2, 2026
Related questions
How can we select language-model populations with low correlated failures when semantic similarity misses generative-process diversity?How can text-conditioned visual autoregressive models increase image diversity without sacrificing quality?How can we compare language models’ conditional behavior and predict the effects of prompt changes?How can text-to-image models preserve variation in unspecified visual factors under long, semantically dense prompts?