Brain-Grounded Axes for Reading and Steering LLM States
Sandro Andric
TL;DR
The paper tackles the challenge of grounding LLM interpretability in external measures by building a brain-derived, MEG-based semantic axis atlas and using a lightweight adapter to read and steer LLM states without fine-tuning. The authors establish a pipeline that derives ICA-based axes from a MEG PLV atlas, validates them against independent lexica, and demonstrates cross-model readout and steering in TinyLlama, Qwen-0.5B, and GPT-2, with a strong emphasis on a lexical-frequency axis that improves fluency under perplexity controls. They further compare brain-derived steering to text-based controls, showing a brain axis can produce more favorable perplexity and robust log-frequency shifts, while also providing cross-subject and cross-model evidence for the axis structure. Exploratory fMRI anchoring suggests potential alignment but remains population-level and sensitive to hemodynamic modeling. Overall, this work presents a neurophysiology-grounded interface that yields interpretable, controllable handles for LLM behavior without modifying the underlying model.
Abstract
Interpretability methods for large language models (LLMs) typically derive directions from textual supervision, which can lack external grounding. We propose using human brain activity not as a training signal but as a coordinate system for reading and steering LLM states. Using the SMN4Lang MEG dataset, we construct a word-level brain atlas of phase-locking value (PLV) patterns and extract latent axes via ICA. We validate axes with independent lexica and NER-based labels (POS/log-frequency used as sanity checks), then train lightweight adapters that map LLM hidden states to these brain axes without fine-tuning the LLM. Steering along the resulting brain-derived directions yields a robust lexical (frequency-linked) axis in a mid TinyLlama layer, surviving perplexity-matched controls, and a brain-vs-text probe comparison shows larger log-frequency shifts (relative to the text probe) with lower perplexity for the brain axis. A function/content axis (axis 13) shows consistent steering in TinyLlama, Qwen2-0.5B, and GPT-2, with PPL-matched text-level corroboration. Layer-4 effects in TinyLlama are large but inconsistent, so we treat them as secondary (Appendix). Axis structure is stable when the atlas is rebuilt without GPT embedding-change features or with word2vec embeddings (|r|=0.64-0.95 across matched axes), reducing circularity concerns. Exploratory fMRI anchoring suggests potential alignment for embedding change and log frequency, but effects are sensitive to hemodynamic modeling assumptions and are treated as population-level evidence only. These results support a new interface: neurophysiology-grounded axes provide interpretable and controllable handles for LLM behavior.