Average percent difference in BERTScore from starting to ending DTG sets is 0.08%

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

• Top-5 instructions by µ(1→2) at ℓ=12 achieve average cosine similarity .9893 and average accuracy .5645 on gsm8k_adv for Gemma3-4B-ITfinding0.726
  High cosine similarity for Gemma3 steering vectors suggests strong linear reflection structure.
• Beam search DTG achieves ending NLI Diversity of 5.35 vs nucleus sampling's higher performance, starting from -5.05finding0.719
  Confirms nucleus sampling produces more semantically diverse outputs than beam search
• VS-Standard achieves KL divergence of 0.54 from pretraining distribution on Open-Ended QA, vs. 3.14 for Direct and 0.58 for Sequencefinding0.708
  Shows VS substantially better approximates the pretraining distribution than baseline methods
• Emergent misalignment appears when between 25% to 75% of the fine-tuning dataset consists of incorrect data (health: 25%, code: 75%)finding0.699
  Dataset mixture experiments establish the fraction of incorrect data needed to induce misalignment
• Typicality bias rates in instruction-tuned models remain at similar or higher levels compared to their base model counterpartsfinding0.698
  Shows typicality bias is preserved through instruction tuning and RLHF, not introduced by alignment
• Gemma-2-27B MT-Bench score slightly decreased from 8.81 to 8.40 ± 0.15 after SOO fine-tuningfinding0.694
  SOO fine-tuning caused a small decrease in Gemma-2-27B general capabilities
• VS-Multi achieves average accuracy of 37.5% on math benchmarks (avg of MATH500, OlympiadBench, Minerva) with Gemini-2.5-Flash as generator and Qwen3-4B as SFT model, vs. 40.7% baselinefinding0.693
  Best VS result in synthetic data generation for math, demonstrating downstream improvement through diversity
• Typicality weight α = 0.65 ± 0.07 (p < 10^-14) on correctness-matched HelpSteer pairs using GLM-4.5 as referencefinding0.693
  Empirical evidence for positive typicality bias consistent across different base model references