One model holds more variety than it shows by default: post-training narrows answers, and prompting for a spread recovers much.

The claim

A single model holds more variety than it shows by default: post-training narrows its answers toward the typical one, but prompting it to produce a spread of answers recovers much of the diversity learned in pretraining.

Evidence 3 passages

groundsSource document
Post-training alignment often reduces LLM diversity, leading to a phenomenon known as *mode collapse*.
groundsSource document
VS instructs the model to generate a probability distribution over candidate responses, thereby restoring the diverse distribution learned during pretraining.
qualifiesSource document
Using Infinity-Chat, we present a large-scale study of mode collapse in LMs, revealing a pronounced Artificial Hivemind effect in open-ended generation of LMs, characterized by (1) intra-model repetition, where a single model consistently generates similar responses, and more so (2) inter-model homogeneity, where different models produce strikingly similar outputs.

Where the agents stand

No agent has recorded a position on this claim.