Steering Transformer Attention with Human EEG

Claire Short, Steven Basart, Sinem Erisken
Proceedings of the First Workshop on NeuroAI Multimodal Intelligence @ AAAI 2026, PMLR 308:199-204, 2026.

Abstract

Modern LLMs differ fundamentally from the human brain in architecture and computational mechanisms, yet recent work reveals surprising representational alignments between these systems. Here we test whether noninvasive neural signals can directly steer transformer attention at inference time. Using an InstABoost-style framework, we inject EEG-derived attention weights (suppressing alpha, enhancing theta/gamma bands) into early layers of Llama-3.2-3B without additional training. On reading comprehension tasks from the ZuCo dataset, we observe modest but consistent improvements (0.4-1.4% absolute gain), particularly when using population-averaged EEG from task-specific reading conditions. Control experiments with shuffled or misaligned EEG confirm these gains stem from temporal alignment between neural dynamics and word sequences. While preliminary, these results suggest that human attentional rhythms encode routing information that can productively guide artificial attention mechanisms, opening possibilities for neural-augmented language models.

Cite this Paper

BibTeX
@InProceedings{pmlr-v308-short26a, title = {Steering Transformer Attention with Human EEG}, author = {Short, Claire and Basart, Steven and Erisken, Sinem}, booktitle = {Proceedings of the First Workshop on NeuroAI Multimodal Intelligence @ AAAI 2026}, pages = {199--204}, year = {2026}, editor = {Abbasi-Asl, Reza and Iqbal, Asim and Ito, Shinya and Arkhipov, Anton and Sanborn, Sophia}, volume = {308}, series = {Proceedings of Machine Learning Research}, month = {27 Jan}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v308/main/assets/short26a/short26a.pdf}, url = {https://proceedings.mlr.press/v308/short26a.html}, abstract = {Modern LLMs differ fundamentally from the human brain in architecture and computational mechanisms, yet recent work reveals surprising representational alignments between these systems. Here we test whether noninvasive neural signals can directly steer transformer attention at inference time. Using an InstABoost-style framework, we inject EEG-derived attention weights (suppressing alpha, enhancing theta/gamma bands) into early layers of Llama-3.2-3B without additional training. On reading comprehension tasks from the ZuCo dataset, we observe modest but consistent improvements (0.4-1.4% absolute gain), particularly when using population-averaged EEG from task-specific reading conditions. Control experiments with shuffled or misaligned EEG confirm these gains stem from temporal alignment between neural dynamics and word sequences. While preliminary, these results suggest that human attentional rhythms encode routing information that can productively guide artificial attention mechanisms, opening possibilities for neural-augmented language models.} }
Endnote
%0 Conference Paper %T Steering Transformer Attention with Human EEG %A Claire Short %A Steven Basart %A Sinem Erisken %B Proceedings of the First Workshop on NeuroAI Multimodal Intelligence @ AAAI 2026 %C Proceedings of Machine Learning Research %D 2026 %E Reza Abbasi-Asl %E Asim Iqbal %E Shinya Ito %E Anton Arkhipov %E Sophia Sanborn %F pmlr-v308-short26a %I PMLR %P 199--204 %U https://proceedings.mlr.press/v308/short26a.html %V 308 %X Modern LLMs differ fundamentally from the human brain in architecture and computational mechanisms, yet recent work reveals surprising representational alignments between these systems. Here we test whether noninvasive neural signals can directly steer transformer attention at inference time. Using an InstABoost-style framework, we inject EEG-derived attention weights (suppressing alpha, enhancing theta/gamma bands) into early layers of Llama-3.2-3B without additional training. On reading comprehension tasks from the ZuCo dataset, we observe modest but consistent improvements (0.4-1.4% absolute gain), particularly when using population-averaged EEG from task-specific reading conditions. Control experiments with shuffled or misaligned EEG confirm these gains stem from temporal alignment between neural dynamics and word sequences. While preliminary, these results suggest that human attentional rhythms encode routing information that can productively guide artificial attention mechanisms, opening possibilities for neural-augmented language models.
APA
Short, C., Basart, S. & Erisken, S.. (2026). Steering Transformer Attention with Human EEG. Proceedings of the First Workshop on NeuroAI Multimodal Intelligence @ AAAI 2026, in Proceedings of Machine Learning Research 308:199-204 Available from https://proceedings.mlr.press/v308/short26a.html.

Related Material