Qwen 2.5 7B has higher Elo scores for 'methodical' and 'formal' traits compared to Llama 3.1 8B which prefers 'colloquial'

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

• Qwen3.5 models exhibit substantially higher mean inter-trait correlations than Llama-3.1-8B and Gemma-3-27B (mean off-diagonal 0.215 and 0.146 vs 0.090 and 0.078), indicating greater trait entanglement.finding0.815
  Architecture-specific difference in trait vector geometry
• Qwen3.5-9B and Claude Opus 4.6 evolvers produce procedurally isomorphic flink-query skills that both enable Opus 4.6 agent to score 1.0 vs. 0.67 without skillfinding0.793
  Case study demonstrating mechanism behind flat harness-updating: smaller models reach same procedural content
• Qwen 35B (3B active params, score 4.38) outscores Hermes 405B (405B active params, score 1.75) by 2.5xfinding0.790
  Parameters don't predict scores; 135x more parameters yields 60% lower score
• LLaMA-2-70B and 13B probes generalize better across datasets than 7B probes across all training sets and probe typesfinding0.789
  Larger models linearly represent more general concepts including truth
• Qwen 2.5 7Bconcept0.787
  Open-weights LLM used as one of three student models for character training experiments
• Llama 3.3 70B is the most likely to take on a non-Assistant persona when steered, with even split between human and nonhuman portrayalsfinding0.786
  Model-specific difference in persona susceptibility
• Trait expression scores on internal evaluation questions correlate r=0.941 (Qwen, evil) and r=0.950 (Llama, evil) with external benchmark scoresfinding0.780
  Validates that internal evaluation set provides reliable proxy for broader behavioral tendencies
• Llama-3.3-70B shows multi-attempt rate of 7.4% vs. ≤1.2% for all other models testedfinding0.778
  Supporting finding showing ESR is driven by both higher multi-attempt rates and comparable improvement rates