ΨLM: Coupling Frozen Language and Physics Models through Trainable Latent Bridges
We ask whether a pretrained language model and a pretrained physics model can run as one system — two frozen networks answering a single question with communication carried entirely by hidden states, the way the brain’s hemispheres cooperate across the corpus callosum. As of August 2026 we find no published work coupling a pretrained LLM to a neural-operator physics model through a bidirectional latent channel at inference. We build such a system, ΨLM, on a consumer laptop (Apple M2, 24 GB), in four steps. (1) Tool-loop baselines reproduce the known capability-gap effect: a simulator adds +28.3 points to a 3B model on physics QA but hurts a 0.5B model. (2) We produce, to our knowledge, the first public reproduction of the Bicameral Model [12]: two frozen Qwen2.5-0.5B streams coupled by a 6.2M-parameter gated hidden-state interface reproduce the paper’s phase transition, reaching 100% exact tool recall through the latent channel alone. (3) With the auxiliary language model replaced by a frozen Fourier Neural Operator, ΨLM answers Burgers-equation field-value questions at 100% (tolerance ±0.05; MAE 0.014) where the LLM alone achieves 5%, matching the accuracy of an oracle that states the answer in text — with no text at the interface. (4) The result survives hardening: held-out question families expose a clean generalization law (support coverage — readouts generalize only where their training support covers the test distribution, a hypothesis we confirm by refuting the alternative), and the architecture transfers to 2D reaction–diffusion with a pretrained physics foundation model (DPOT-Tiny) as the hemisphere, scoring 95%, within five points of the oracle ceiling. All experiments, including two failed designs and one refuted hypothesis, are reproducible from the public repository on a single consumer machine.