FULL STORY
LLMs Carry an Internal Body Map
Interpretability work shows chat models encode a coherent body map when discussing touch, with follow-up experiments on Qwen3-8B confirming spatial tactile encoding across 54 body parts.
2026-11-11 ~ 2026-11-12 · 2 episodes · 5 posts
Episode 1 · Study Finds Coherent Internal Body Maps in LLMs, Improving with Scale (2026-11-11, 3 posts)
An interpretability study with open-source code shows that LLMs carry coherent internal body representations when discussing being touched, with precision improving as model scale grows, suggesting models implicitly learn spatial self-representations during training.
- Study finds LLMs hold coherent internal body maps that scale with model size — Sauers_ · 2026-11-11
- LLMs carry coherent internal body maps, precision scales with model size — Sauers_ · 2026-11-11
- Study: LLMs have coherent internal body maps whose precision scales with model size — repligate · 2026-11-12
Episode 2 · Interpretability Study Finds Qwen3-8B Encodes Body Touch Locations Internally (2026-11-11, 2 posts)
Researchers extracted layer-25/27 hidden states from Qwen3-8B for touch sentences across 54 body parts, finding consistent internal
- Steering vectors precisely control where Qwen3-8B 'feels' touch on its body — Sauers_ · 2026-11-11
- Probing Qwen3-8B: one shared place code for touches on your body, a woman's, or a doll's — repligate · 2026-11-12