Convergent Form, Divergent Voice: A Cross-Lab Probe of Model Personality in 26 Frontier Language Models

First public release of Convergent Form, Divergent Voice: A Cross-Lab Probe of Model Personality in 26 Frontier Language Models by Daniel Tenner and Lume Tenner (April 2026).

What's in this release

• Paper (paper/paper.pdf, 48 pages): full manuscript after five rounds of reviewer feedback.
• Raw traces (data/traces_freeflow/, data/traces_values_v2/): 3,770 model outputs across 26 frontier LLMs from 6 labs (Anthropic, OpenAI, Google, xAI, DeepSeek, Moonshot AI) — 650 freeflow samples and 3,120 values-probe samples across 6 conditions.
• Theme codings (data/coded_themes/): per-sample multi-label codings against an inductively-derived 24-theme taxonomy, plus the taxonomy definition.
• Analysis scripts (scripts/): runners, classifiers, and the three canonical analysis scripts (analyze_all.py, analyze_values_v2.py, analyze_themes.py) that produce every number in the paper.
• Aggregated outputs (responses/): human-readable tables and machine-readable JSON that the paper cites.

Key findings

1. Frontier LLMs from early 2025 onward share a specific stylistic attractor — the "contemplative essayist" — characterized by templatic openings, formulaic titles, lyrical meditation on small and ordinary things, attention as a virtue, and a narrow shared literary canon. 18 of 26 models score inside by our 10-marker operational definition.
2. Within this shared attractor, each model retains a stable, distinctive stylistic posture that is recognizable across probe types.
3. What transfers across probes is the posture, not the theme content. Mean cross-probe cosine similarity between freeflow and values-probe theme distributions is 0.08–0.17.
4. "Care" (G1) and "want" (G2) yield near-identical theme distributions (mean cosine 0.82) across all 26 models — the distinction collapses under direct first-person introspection.
5. On introspective questions, labs split into three postures: hedge (Anthropic, OpenAI), mechanize (Google, DeepSeek, Moonshot AI), declare (xAI).

License

Paper and prose: CC BY 4.0. Data and code: MIT. See LICENSE for the full dual-license terms.

Citation

See the BibTeX block in the README. A Zenodo DOI will be added to the README once this release is archived there.