Stanford reveals cause of model conformity, boosts diversity 2.1x with multi-option prompts
simplifyinAI · x · 2026-08-23
Stanford researchers found that AI models' tendency to produce repetitive answers stems from human labelers preferring more "familiar" responses during RLHF, baking this bias into the model. Instead of retraining, simply asking the model to generate multiple answers with probabilities (e.g., "Generate 5 jokes and their corresponding probabilities") significantly mitigates this issue. Tests show this method increases diversity in creative writing (poems, stories, jokes) by 1.6-2.1x, with no drop in factual accuracy or safety, and more capable models benefit more.
Related event: Generating Multiple Options Boosts LLM Answer Diversity, Study Finds(3 posts)→
More from Models
- Free account UI shows the top-tier model while serving a cheaper one — Sure_Raccoon5330 · 2026-08-23
- Can Local Qwen Models Handle Real-World System Programming like GTK4/Qt? — MongoWithBongoss · 2026-08-23
- Ox-Alpha rumored to be GLM 5.3 Flash: open, cheap, top-tier — daniel_mac8 · 2026-08-23
- Kimi K3 Claims Private Reasoning is Hidden Despite Being Visible — eliebakouch · 2026-08-23
- Analysis Suggests Ox Alpha Distilled from GLM-5 Weights — koltregaskes · 2026-08-23
- Users Praise Gemini 3.7 Flash for Its Exceptional Speed — MeganRisdal · 2026-08-23