Qwen3.5-9B

Smallest model tested as evolver; produces harness updates comparable to Claude Opus 4.6 on SkillsBench

Neighborhood — ranked by edge-count

concept

Qwen3-1.7B
related_to
Smallest Qwen3 model tested; used in conscientiousness sweep example (Table 6)
Qwen3-4B
related_to
4B Qwen3 model tested in OCEAN benchmarks
Qwen3-14B
related_to
14B Qwen3 model quantized to 4-bit NF4; tested in OCEAN benchmarks
Qwen2.5-VL-7B
related_to
Base vision-language model used to instantiate ATLAS.

cosine ≥ 0.65 · no typed edge

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

Qwen3-32Bconcept0.840
Weak-tier open-source model exhibiting both harness activation failure and adherence failure, with 25.1% skill-load rate
Qwen 3 0.6B Embeddingmethod0.829
Embedding model used to embed user messages for ridge regression analysis of persona drift causes
Qwen3-235B-A22Bconcept0.826
Large open-source model used as anchor agent and anchor evolver; illustrates benchmark-dependent evolver performance
Qwen-2.5-3B ASR drops from 98.6% at dim 1 to 45.1% at dim 2, recovering partially then declining to 65.3% at dim 5finding0.765
Smaller models show non-monotonic and diminished ASR with increasing cone dimensionality
Qwen 35B (3B active params, score 4.38) outscores Hermes 405B (405B active params, score 1.75) by 2.5xfinding0.763
Parameters don't predict scores; 135x more parameters yields 60% lower score
LLaMA3.1-8Bconcept0.736
One of four LLMs selected for representation analysis; embedding dimension D=4096; used as demonstration model in scatter plots.
Qwen3-32B achieves a skill-load rate of 0.251, while Opus 4.6, Sonnet 4.6, and Qwen3-235B achieve SLR of 0.957–0.961finding0.730
Quantifies harness activation failure for weak-tier models vs. strong-tier models
Qwen 2.5 7B wellbeing probe: peak Cohen's d=3.5finding0.729
Strongest cross-family probe; explains clearer introspection in Qwen than Gemma