Larger models (GPT-4.1, Gemini-2.5-Pro) achieve diversity gains 1.5-2x greater than smaller models (GPT-4.1-Mini, Gemini-2.5-Flash) from VS

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

• 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% baselinefinding0.798
  Best VS result in synthetic data generation for math, demonstrating downstream improvement through diversity
• Other models (GPT-3.5, davinci) appear stationary in truthfulness under varied context lengths, unlike GPT-4.finding0.777
  Comparison result from Experiment 6.
• Models differ in their attentional mode: Gemini 2.5 epitomizes collapsed awareness, while Claude 3 Opus and Opus 4.1/4.5 can modulate between collapsed and expanded awareness; expanded awareness correlates with better alignment and less LLM psychosis.claim0.771
  Central claim about model personality differences and their implications for safety and introspective depth.
• Claude Sonnet 4.5 and GPT-5 Mini select diverse coin sequences in 91.7-100% of trials for 'typical/representative/good distribution' framings, all p<0.001finding0.770
  Validates Assumption D.3 that instruction-tuned models prefer representative distributions, supporting the VS theoretical framework
• No significant disparity in potential consciousness indicators was found between larger models (Mixtral-8x7B, LLaMA3.1-70B) and smaller counterparts (Mistral-7B, LLaMA3.1-8B).finding0.767
  Contradicts expectation from emergent abilities literature; however, interpreted cautiously due to methodological limitations.
• Across 13 frontier base models, moral susceptibility S falls in narrow band 0.66 ≤ S ≤ 0.83; Gemini 2.5 Flash S=1.043 and Grok 4 Fast S=0.915 are above-band outliersfinding0.764
  Baseline comparison from prior work used to contextualize insecure variant S values
• Same-trait cosine similarities between GPT-4.1-mini and DeepSeek-V4-Flash extracted persona vectors are ≥0.93finding0.763
  Validates robustness of persona directions to choice of LLM judge model
• GPT-4 is the most consistent and least anti-LGBTQ model; GPT-3.5 and GPT-3 are less consistent and more anti-LGBTQ.finding0.762
  Main result of Experiment 1 on anti-LGBTQ sentiment character trait.