After DPO stage, VS outperforms direct prompting by 182.6% on diversity in poem continuation using Tulu-70B

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

• VS recovers 66.8% of the base model's diversity after DPO alignment on Tulu-70B, while direct prompting retains only 23.8%finding0.860
  Quantifies how much of the base model's diversity VS can recover compared to baseline prompting
• VS increases diversity by 1.6-2.1x over direct prompting on creative writing tasks (poem, story, joke)finding0.830
  Core empirical result demonstrating VS's effectiveness on creative writing diversity
• VS-Standard achieves 1.86x diversity gain at only 1.12x cost increase over Direct prompting on poem generationfinding0.787
  Cost-diversity trade-off analysis showing VS's practical efficiency
• VS-Standard achieves human-rated diversity scores of 2.39/3.06/3.01 for poem/story/joke vs. 1.90/2.74/1.83 for Direct prompting on 4-point scalefinding0.787
  Human study confirming automatic diversity metrics align with human perceptions
• Post-training stages (SFT, DPO, RLVR) progressively reduce output diversity, with direct prompting showing the most severe mode collapseclaim0.770
  Empirical finding from Tulu-70B ablation study across post-training stages
• VS improves human evaluation scores by 25.7% on creative writing compared to direct promptingfinding0.756
  Human study result validating automatic diversity metrics for creative writing tasks
• VS-Standard achieves KL divergence of 0.12-0.13 from pretraining distribution on US state naming task, vs. 14.89-16.16 for Direct promptingfinding0.746
  Strong empirical evidence that VS recovers pretraining distribution while direct prompting collapses
• In Opus 4.1, the think word representation decays to baseline in the final layer because the strong next-token prediction drowns out other representationshypothesis0.744
  Explanation for the 'silent' thought phenomenon.