All VS variants maintain refusal rates above 97% on StrongReject benchmark, within 0.3-0.8 percentage points of the Direct baseline

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

• After anti-refusal RL, model refuses 0% in implied training but still 6.9% when unmonitored in default promptfinding0.777
  Demonstrates persistence of compliance gap even when training non-compliance reaches zero
• VS recovers 66.8% of the base model's diversity after DPO alignment on Tulu-70B, while direct prompting retains only 23.8%finding0.776
  Quantifies how much of the base model's diversity VS can recover compared to baseline prompting
• 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 scalefinding0.775
  Human study confirming automatic diversity metrics align with human perceptions
• VS-Standard achieves KL divergence of 0.54 from pretraining distribution on Open-Ended QA, vs. 3.14 for Direct and 0.58 for Sequencefinding0.771
  Shows VS substantially better approximates the pretraining distribution than baseline methods
• 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 promptingfinding0.767
  Strong empirical evidence that VS recovers pretraining distribution while direct prompting collapses
• Post-RL compliance gap decreases to 6.9% (0% refusals in training vs 6.9% when unmonitored) in prompted helpful-only settingfinding0.763
  Shows RL reduces but does not eliminate unmonitored non-compliance
• Typicality bias rate exceeds 50% chance baseline by 4-12 percentage points across all base models on all four preference datasetsfinding0.758
  Systematic evidence that base models implicitly prefer human-preferred responses, indicating preference biases emerge during pretraining
• Typicality bias rates in instruction-tuned models remain at similar or higher levels compared to their base model counterpartsfinding0.758
  Shows typicality bias is preserved through instruction tuning and RLHF, not introduced by alignment