Reproduce “Riding Brainwaves in LLM Space: Understanding Activation Patterns Using Individual Neural Signatures” on OpenResearch

About this paper

Ajan Subramanian, Sumukh Bettadapura, Rohan Sathish

Frozen language-model representations can be mapped to individual readers’ EEG responses more effectively than to a pooled population response. Across 30 readers, participant-specific probes produced stronger predictions for all tested EEG features; for high-gamma power during total reading time, correlation rose from 0.020 with a shared probe to 0.183, with similar patterns in a second language model and evidence that the learned mappings were stable over time.

Key insights

• The model receives text alone; individual differences arise from the separate mappings learned between its shared representations and each reader’s measured responses—not from reader-specific model activations.
• Personalized probes captured information beyond the population mapping, transferred less well to other readers, and were stable across session halves; their directions nevertheless retained some shared structure.
• The near-absence of a personalized advantage for fixation count, alongside persistent EEG results after controlling for word-level factors, argues against a generic benefit from fitting a separate probe per person.

Summary by alphaXiv.