Which Attention Heads are like the Human Head? Not the Ones that Compute
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 → 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.
Related
- Behavioral signatures suggest distinct modes of suppressing irrelevant information during tactile temporal attention in human participants
- Simulated field potentials capture multiple simultaneous processes underlying human temporal attention
- Graph attention spatio-temporal graph neural network based human–light interaction modeling for indoor lighting environments
- Attention warps visual-semantic representations across the human cerebellum
- Optogenetic modeling of wake-like transcriptional progression in human SH-SY5Y neuronal cells
BibTeX
@misc{pinier2026which,
title = {{Which Attention Heads are like the Human Head? Not the Ones that Compute}},
author = {Christopher Pinier and Gustaw Opiełka and Hannes Rosenbusch and Taylor Webb and Michael D. Nunez and Claire E. Stevenson},
howpublished = {arXiv},
year = {2026},
month = sep,
doi = {10.48550/arxiv.2609.37991},
eprint = {2609.37991},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2609.37991},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - Which Attention Heads are like the Human Head? Not the Ones that Compute
AU - Christopher Pinier
AU - Gustaw Opiełka
AU - Hannes Rosenbusch
AU - Taylor Webb
AU - Michael D. Nunez
AU - Claire E. Stevenson
PY - 2026
DA - 2026/09/29
PB - arXiv
DO - 10.48550/arxiv.2609.37991
UR - https://arxiv.org/abs/2609.37991
AB - 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 → 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.
ER -