Probing Qwen3-8B: one shared place code for touches on your body, a woman's, or a doll's
repligate · x · 2026-11-12
An interpretability experiment on Qwen3-8B averaged the model's layer-27 (of 36, 4096-dim) hidden states over touch sentences covering 54 body places, then classified new sentences by nearest fingerprint direction. Touches on "your" body landed at the right place 95% of the time, on a woman's body 92%, on a doll's 91% (chance: 2%). The model thus uses one shared place code; the doll's state vectors are shorter (0.65x length) but point the same way — same direction, lower magnitude.
More from Research
- Microsoft Research's ATLAS & JANUS: Neural Samplers for Metallic Glass Materials Discovery — LesterMackey · 2026-11-12
- Steering Qwen3-32B to feel "small" makes it more prone to social engineering — Sauers_ · 2026-11-12
- Why is there no pretrain-postrain co-design? Companies ship their org charts — SonglinYang4 · 2026-11-12
- Google Research trains medical models that understand patient context with clinicians — burny_tech · 2026-11-12
- Every Curve Is a Compressed Argument: Key Data Science Plots Explained — burny_tech · 2026-11-12
- AI formalizes Fermat's Last Theorem in 30 hours, Yang-Mills construction wins Nobel laureate committee sign-off — rand_longevity · 2026-11-12