Linear Digressions Chasing Away Repetitive LLM Responses with Verbalized Sampling
22 snips
Feb 23, 2026 They demonstrate mode collapse with live prompts and show why LLMs fall into repetitive outputs. The episode explains how alignment and annotator bias push models toward typical, conservative replies. It introduces verbalized sampling — asking for multiple responses with probabilities — and reviews experiments showing restored diversity, especially in larger models.
AI Snips
Chapters
Transcript
Episode notes
ChatGPT Stuck On A Few States
- The host demos ChatGPT repeatedly naming US states and it cycles through Oregon, Colorado, and occasionally Texas instead of showing broad variety.
- This live example illustrates mode collapse vividly by contrasting repeated single-answer prompts with expected real-world diversity.
Typicality Bias Drives Mode Collapse
- The episode reframes mode collapse as not solely an algorithmic alignment failure but potentially arising from biased preference data used during alignment.
- Human annotators favor typical, familiar outputs, which sharpens the post-alignment distribution toward safe, repetitive responses.
Preference Data Mirrors Pretraining But Skews Typical
- The paper compares preference datasets used for alignment with pretraining distributions and finds annotator preferences skew toward typical items, mathematically enabling collapse.
- This shows the bias enters via alignment data aggregation, not just the alignment algorithm itself.