Brain aligned attention heads differ from model critical heads
Which Attention Heads are like the Human Head? Not the Ones that Compute
Artificial Intelligence
Summary
The authors studied how certain parts of language models, called attention heads, relate to human brain activity during a simple pattern task. They found that attention heads that look like human brain responses don't always do the main work the model uses to solve the task. Removing those brain-aligned heads harms performance less than removing other important heads identified by analyzing the model's decision process. This suggests that similarity to human brain activity shows how the model reads inputs, but not exactly how it thinks or computes answers.
What this means in practice
- •For ai model developers: Identify which model components are essential for task performance versus those that resemble human brain signals for better model understanding.
- •For neurotechnology engineers: Use insights into brain-like but non-critical model parts to design brain-computer interface components that reflect perception rather than decision logic.
Authors
Christopher Pinier, Gustaw Opiełka, Hannes Rosenbusch, Taylor Webb, Michael D. Nunez, Claire E. Stevenson
Abstract
Brain-AI alignment is often interpreted as a sign that model and brain perform similar computations. Whether the aligned units are causally involved in model computation is rarely checked. On an abstract pattern-completion task (AAABAAA $\rightarrow$ B), we compare LLM attention-head representations with human EEG and test how ablating those heads affects task performance. Alignment and causation dissociate: brain-aligned heads contribute to performance, but their removal is substantially less disruptive than removal of heads selected via attribution patching. We compare two head sets that prior interpretability work defines without reference to the brain: concept vectors (CVs), which represent abstract patterns across formats, and function vectors (FVs), selected for their contribution to correct-answer prediction. Brain alignment shows little association with FV scores, while its association with CV scores varies across models. Among brain-aligned heads, we find recurring attention profiles: one emphasizes distinctive elements (novelty heads), the other repeating elements (repetition heads). The novelty family tracks salience and attends to the same elements that humans look at, yet its removal is less damaging than random ablation on average. Repetition heads contribute modestly to performance and are associated with abstract-pattern representation (CVs). Across 17 models spanning 3B-72B parameters, FV-ranked removal is substantially more disruptive than brain-ranked removal. Brain alignment thus captures how the model reads the stimulus, and only faintly captures how it represents the pattern and solves the task.