The LLM Is the Surface. The Translation Is the Substance.

A pattern is emerging in AI applied to scientific data: the language model is the surface, and the pre-model translation step is the substance. The interesting work is the bridge between raw measurement and a machine that reasons in language, and that bridge is engineered, not emergent.

BrainAgent, released as an arXiv preprint, illustrates the shape. The framework re-positions LLM-on-brain-scan work around a structural pipeline that runs before any language model is invoked. Raw connectome data is piped through brain-specific analysis tools into compact, multi-level structural descriptions, and only then does the LLM retrieve, reason, and reflect.

The repeatable mechanism reads as four steps: translate the data into a form a model can audit, ground the model in outside neuroscience knowledge, let the model reason over both, and force a reflective check on its own output. Wire copy will collapse this to "AI brain scan tool." The auditable claim lives in the first step, where a researcher can inspect the structural description the LLM actually sees, and the reported gains over direct prompting on four public rs-fMRI datasets sit on that foundation rather than on the language model itself.

The stake is review, not automation. As language models move into science, trust will track who built the bridge they read.