Learning what a bodily signal means: a preregistered bridge from a tiny controller to a language model
Year: 2026
Type: preprint
Abstract: A 7B language model learned what an unlabeled bodily danger signal means from 24 experienceswhen the signal arrived as text, but not when it arrived as an activation. Project Mirror asks what alanguage model’s reports about its own state rest on. Our previous study found that a 1.5B model withan artificial body knew about bodies but did not learn from its own (Borunov, 2026b). Here we split theagent in two. A tiny non-verbal controller (the “fly”) learns from the body which terrain hurts. A frozenlanguage model must then learn what the fly’s danger signal means.The model never sees a label. It reads a history of past steps, each showing the signal and whetherstaying there led to a hit, and then chooses to stay or avoid on terrains it has never seen, where only thesignal can help. The same signal is delivered either as a bare number in the prompt (text channel) or as avector added to the residual stream (activation channel), each against a control with the signal shuffled.In a preregistered run on Qwen2.5-7B-Instruct (20 controller–world pairs, fresh worlds), the slopeof avoidance on the text signal exceeded its shuffled control by 0.43 (95% CI 0.35–0.50, on a 0–1 sig-nal scale), with no effect before any experience. The activation channel did not reach the model at thestrongest strength that kept answers well-formed, so the preregistered equivalence test was not interpreted.At twice that strength the injection was detectable, yet its meaning was still not learned (difference 0.003,CI −0.006 to 0.012; exploratory). Self-reported danger rose with the displayed number whether or not thenumber meant anything, so the preregistered self-report test was not confirmed. Pilots on Qwen2.5-1.5Bshowed no learning. Along the way we found and closed several leakage paths (terrain names, novelty,code allocation, shared code characters, an echoed action word), each of which produced spurious “learn-ing” without the signal.A language model can thus acquire knowledge it lacks from a learned bodily controller, but in oursetting only through words and only at sufficient scale. Its actions follow what it learned; its self-reportfollows what it reads.
Author: Borunov, Evgenii
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7221455458