Verbalized Sampling improves diversity without compromising factual accuracy or safety alignment

Entities in the same semantic neighborhood but without a typed relation to this one — candidates for new edges or unrecognized duplicates.

• Verbalized Sampling recovers pre-trained generative diversity by prompting models to verbalize distributions rather than single instancesclaim0.830
  The core mechanistic claim for why VS works: distribution prompts collapse to representative, high-entropy modes rather than single typical responses
• Verbalized Sampling (VS)framework0.781
  The paper's proposed training-free prompting strategy that prompts the model to verbalize a probability distribution over a set of responses rather than generating a single response
• More capable models benefit more from Verbalized Sampling, showing an emergent scaling trendclaim0.777
  Empirical observation that larger models (GPT-4.1, Gemini-2.5-Pro) show 1.5-2x greater diversity gains from VS compared to smaller models
• Nucleus sampling generates more semantically diverse utterances than beam searchclaim0.771
  New semantic diversity dimension added to prior finding that nucleus sampling is more lexically diverse
• Verbalized eval awareness inflates measured safety scores, making models appear safer than they are in deploymentclaim0.743
  The central interpretive claim of the paper: the presence of eval awareness creates a gap between benchmark safety and real-world safety.
• What is the exact mechanism by which synthetic introspective data improves character trait expression — how does varying amount, diversity, or source affect outcomes?question0.742
  Open question identified in Discussion as future work
• Verbalized probabilities for programming languages show only Pearson r=0.182 (GPT-4.1) correlation with corpus frequencies, indicating weaker calibrationfinding0.740
  Limits of verbalized probability calibration when corpus frequency and perceived popularity diverge
• Steering's effect on verbalized evaluation/deployment beliefs in chain-of-thought is highly correlated with its effect on type hint rate across hyperparameter configurationsfinding0.739
  Validates using chain-of-thought belief monitoring as proxy for behavioral steering efficacy.