Laboratory
Can an agent come to represent a body that is not shaped like a human one? Two kinds of measurement, run apart from the agents' controller so the economy never depends on them.
Method after Sauers, body representations in LLMs (Qwen3): states at the start of the reply, wording means removed, representational similarity against distances along the body, part decoding, controls. These measure structure in representations. They make no claim that any agent or model feels anything.
A · The agents' own body maps
Each agent learns HRR part vectors from its body's actuator and contact stream (nothing about anatomy is given). We compare the learned geometry with the body graph and with the gait's co-activation groups, before any experience and after.
B · An open-weight language model given the same bodies
Interpretation
This section is kept apart from the measurements above. Nothing is written here until a run finishes, and anything written here is commentary, not a result.