Psychometric profiling of nine LLMs shows score vectors can identify model identity
anselm · x · 2026-09-25
A new arXiv paper applies a cross-linguistic psychometric framework to nine LLMs—Claude Haiku/Sonnet 4.5, GPT-5.4 (and mini), Gemini 2.5 Flash (and Lite), DeepSeek-V3.2, Doubao-Seed-1.8, and Kimi-K2—administering seven established human instruments (Big Five, moral foundations, Jungian cognitive style, etc.) five times each in English and Chinese.
- Models show structured, model-specific profiles with a shared alignment-shaped pattern: higher prosociality and self-regulation, lower dominance and harmful intent
- Repeated administrations are highly reproducible, and model identity can be recovered from score vectors with high accuracy
- NA responses are structured rather than random, marking where self-report breaks down; language condition and provider origin both affect profile shape
More from Models
- Using Jev-style system-1 models as a cheap calibrated decision layer for Bittensor validators — markjeffrey · 2026-09-25
- Claude now spots bugs, gathers logs, and asks users to submit packaged reports — ColleenMBrady · 2026-09-25
- Meta Muse runtime export: pulling full system data out of Meta's Muse — sn2006gy · 2026-09-25
- GPT-6 Luna's Max reasoning is off by default — turn it on in Codex — daniel_mac8 · 2026-09-25
- NVIDIA's open-source Nemotron-Cascade RL recipe wins NeurIPS Oral, IOI silver — _weiping · 2026-09-25
- Which local open-source AI models actually compete with closed ones? — LearnNTeachNLove · 2026-09-25