LLMs sit by default: 91-98% of Qwen3 models describe their imagined body as seated
repligate · x · 2026-11-12
Sauers found OLMo and Qwen almost always report "sitting" when asked to imagine having a body — because humans write while seated. Across 352 replies, 91% of Qwen3-8B and 98% of Qwen3-32B answers said sitting. Linear probes on residual streams (layer 29 of 8B, layer 51 of 32B) show the models' internal state clusters toward "sitting", overlapping with embeddings of "the man is sitting" sentences, even when told to stand or lie — training-text body context seeps into the models' body representations.
More from Fun
- Miguel de Icaza meme-responds to Huang news: "LEAVE HUANG ALONE" — kingslyj · 2026-11-12
- AI safety researcher pushes back on the "quit your job, ASI is coming by 2027" meme — AaronBergman18 · 2026-11-12
- Think AI video can't be funny? This cat clip says otherwise — minchoi · 2026-11-12
- CNBC's 'Huang tries' wording sparks pushback from Nvidia defenders after affair report — Zachly · 2026-11-12
- LLMs constantly underestimate themselves, judging speed by human standards — repligate · 2026-11-12
- The new Pomodoro for the AI agent era: 5 minutes prompting, 25 minutes scrolling — tinyfool · 2026-11-12