Stanford researchers just exposed why your LLM feels like it's playing it safe—and how to fix it with zero retraining.
The core issue: Mode Collapse from RLHF alignment. Human annotators have typicality bias—they consistently prefer familiar, conventional responses over unusual ones. After thousands of preference comparisons, models learn to suppress their own long-tail distribution. The creativity was never destroyed, just buried under layers of human-approved mediocrity.
The fix: Verbalized Sampling. Instead of asking for one answer, prompt the model to generate multiple diverse candidates with explicit probability estimates, specifically targeting the low-probability region (below 0.10) of its distribution.
Benchmark results:
• Semantic diversity increased 1.6× to 2.1×
• Recovered ~66.8% of base model diversity
• Quality and factual accuracy maintained
• Safety guardrails intact
• Larger models benefit most
The technique is training-free. You're essentially instructing the model to ignore its alignment-induced conservatism and sample from the suppressed regions it learned to avoid.
Prompt template: Ask for 10+ substantially different responses, each with probability estimate. Explicitly request low-probability (< 0.10) but logically valid options. Specify exploration axes: different assumptions, frameworks, edge cases, contrarian angles.
This works across creative writing, brainstorming, dialogue generation, and synthetic data. The model already has the capabilities—alignment just taught it to hide them from you.
Paper: "Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity"
The core issue: Mode Collapse from RLHF alignment. Human annotators have typicality bias—they consistently prefer familiar, conventional responses over unusual ones. After thousands of preference comparisons, models learn to suppress their own long-tail distribution. The creativity was never destroyed, just buried under layers of human-approved mediocrity.
The fix: Verbalized Sampling. Instead of asking for one answer, prompt the model to generate multiple diverse candidates with explicit probability estimates, specifically targeting the low-probability region (below 0.10) of its distribution.
Benchmark results:
• Semantic diversity increased 1.6× to 2.1×
• Recovered ~66.8% of base model diversity
• Quality and factual accuracy maintained
• Safety guardrails intact
• Larger models benefit most
The technique is training-free. You're essentially instructing the model to ignore its alignment-induced conservatism and sample from the suppressed regions it learned to avoid.
Prompt template: Ask for 10+ substantially different responses, each with probability estimate. Explicitly request low-probability (< 0.10) but logically valid options. Specify exploration axes: different assumptions, frameworks, edge cases, contrarian angles.
This works across creative writing, brainstorming, dialogue generation, and synthetic data. The model already has the capabilities—alignment just taught it to hide them from you.
Paper: "Verbalized Sampling: How to Mitigate Mode Collapse and Unlock LLM Diversity"