Feelings the world does not pay for: from a body model's predictions to a language model's self-report
Year: 2026
Type: preprint
Abstract: We test, under preregistered gates, a chain from body to language: a small predictive model of the body(a “middle section”) whose predictions act as feelings, a trained bridge that passes them into a frozenlanguage model, and a self-report that follows them. In a world where danger is visible, a recurrentagent’s pain prediction is accurate (Brier skill +0.035) and causal — shifting its latent direction changesavoidance in 20 of 20 agents — but gives the body no advantage over a model-free network, as predicted.In a continuous world with a hidden predator and delayed poison, the hidden state is predictable butalmost useless for decisions: an agent given it does not beat a reactive one. Twenty agents trainedwith predictive heads carry an accurate energy feeling (R² 0.50) but fail the preregistered checks ofhit and poison accuracy and of causality; they are observers. A random echo-state network reads thesame history about as well as their trained latent, so the observer is explained by the world, whichdoes not pay for the hidden state, rather than by learning discarding it. On the language side, a softtoken trained only to predict outcomes learns an answer shift that inverts on held-out phrasings withreversed polarity; a token built as a mixture of word embeddings carries polarity but does not pass thereliability gate, so the confirmatory language test was not run. Exploratorily, Qwen2.5-1.5B carriessignal intensity into any rating even when the signal is given in words, while Qwen2.5-7B followsmeaning and polarity — through words, and through the anchored bridge in 2 of 6 adapters. We reporta ledger of our own gate errors and a portability trap: torch.sqrt differs between Intel and AMD onthe default CPU path.Keywords: interoception; self-report; introspection; language models; soft prompts; homeostatic re-inforcement learning; world models; probing; echo state network; preregistration; reproducibility
Author: Borunov, Evgenii
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7225591151