Psycholinguistic analysis finds Claude Opus 5 the most out-of-distribution model tested
gleech · x · 2026-10-24
A study measuring how far model outputs deviate from baseline human language distributions, using psycholinguistic features and a round-trip paradigm, finds clear differences across models. Claude Opus 5 stands out as the most out-of-distribution across Chat, Coding, and Essays, well above other Claude versions, while Kimi, GLM, Llama, and Olmo variants deviate far less; base models show the smallest shifts.
More from Models
- Codex 'dumbed down'? User says it skipped web search and re-engineered an app instead — mark_k · 2026-10-24
- Dessa's dialog-only voice training isn't new: dGSLM did it in 2022 with just 2k hours — rdesh26 · 2026-10-24
- Google unexpectedly refunded a user's AI Pro subscription payment — Think-Caterpillar36 · 2026-10-24
- Dev compares Claude vs OpenAI's Astra: Claude fixes its own mistakes, Astra spins for days — StewartalsopIII · 2026-10-24
- Qwen 27B local RAG on a 3090 ties Opus 5.5 on Korean benchmark — Front-University4363 · 2026-10-24
- Observer: Moonshot's Compute Instincts Belong at Google, DeepSeek's Rookies Built the Most Efficient LLM — teortaxesTex · 2026-10-24