Probing Qwen3-8B: Action Embeddings Land on the Right Body Parts, Up to 100%
Sauers_ · x · 2026-11-12
A probe of Qwen3-8B's embedding space shows action sentences like "You kick the ball" map nearest to the body part performing the action (e.g. 78% for kick vs 63% word-similarity baseline across 23 body parts). In reverse tests, adding an action's vector (wink/punch/lift) makes the model report matching bodily sensations — "a subtle flicker in my left eye" (94%), up to 100% — suggesting LLM representations spontaneously encode an embodied spatial structure.
More from Research
- 1920 Popular Mechanics clipping shows eclipse that confirmed light bends under gravity — generativist · 2026-11-12
- Speculative decoding drafter benchmark: AngelSpec DFly beats EAGLE-3 variants — maxsloef · 2026-11-12
- ELIZA's Lost Code: 2,600 Lines Found in MIT Archives in 2021, Retyped by Hand — demian_ai · 2026-11-12
- 1.5-Year Study: Cultural Bias in AI Curiosity and Its Role in Multi-Agent Collaboration — radamihalcea · 2026-11-12
- ContactBench 1.0: A Benchmark for Contact-Rich Robot Manipulation With Printable Assets — leto__jean · 2026-11-12
- Formalizing all of nLab joked as the 'final boss' of AI math — burny_tech · 2026-11-12