cs.CLMar 23, 2026

Riding Brainwaves in LLM Space: Understanding Activation Patterns Using Individual Neural Signatures

Authors: Ajan Subramanian, Sumukh Bettadapura, Rohan Sathish

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Consumer-grade EEG is entering everyday devices, from earbuds to headbands, raising the question of whether language models can be adapted to individual neural responses. We test this by asking whether frozen LLM representations encode person-specific EEG signals, directions in activation space that predict one person's brain activity but not another's. Using word-level EEG from 30 participants reading naturalistic sentences (ZuCo corpus), we train a separate linear probe for each person, mapping hidden states from a frozen Qwen 2.5 7B to that individual's EEG power. Person-specific probes outperform a single population probe on every EEG feature tested; for high-gamma power, the person-specific probe achieves rho = 0.183, a ninefold improvement over the population probe (rho = 0.020, p < 10^-4). A negative control, fixation count, shows no person-specific advantage (p = 0.360); fixation count reflects word length and frequency rather than individual cognition. The individual directions are temporally stable (split-half cosine = 0.824), non-transferable across people (self rho = 0.369 vs. other rho = 0.143, p < 10^-19), and distinct from the shared population signal: person-specific probes retain predictive power after the population component is removed. The person-specific signal concentrates in the model's deep layers, rising consistently with depth and peaking at Layer 24 of 28. The results are consistent across architectures (LLaMA 3.1 8B) and survive word-level confound controls. Frozen language models contain stable, person-specific neural directions in their deep layers, providing a geometric foundation for EEG-driven personalization.

Figures & tables

Figure 1: (a) Different readers produce distinct neural signatures for the same word; each person-specific probe learns a different direction in the LM’s activation space (“pop” = population). (b) Probing pipeline: words pass through a frozen LM and PCA; 30 person-specific Ridge probes are compared against a single population probe. The person-specific probe achieves ρˉ=0.183\bar{\rho}=0.183 vs. 0.0200.020 for the population probe on TRT_g2 (high- γ\gamma , Layer 24).
FeaturePerson ρ\rhoPop ρ\rhoΔ\Deltapp
meanPupilSize0.2610.071+0.189<10−4<10^{-4}
TRT_g2 (gamma high)0.1830.020+0.163<10−4<10^{-4}
TRT_g1 (gamma low)0.1560.043+0.113<10−4<10^{-4}
TRT_b2 (beta high)0.1450.069+0.077<10−4<10^{-4}
TRT_t1 (theta low)0.1000.024+0.075<10−4<10^{-4}
TRT_a1 (alpha low)0.1270.054+0.073<10−4<10^{-4}
Table 1: Person-specific vs. population probe performance (Spearman ρ\rho ) at Layer 24, PCA 50. All EEG comparisons are significant at p<10−4p<10^{-4} (paired tt -test, N=30N=30 ). nFixations (a text-structural measure) shows no person-specific advantage ( p=0.360p=0.360 ).
Figure 2: Layer-wise probe performance on Qwen 2.5 7B (ZuCo, TRT_g2). Person-specific probes (solid blue, mean ρ\rho across 30 participants) increase with depth, peaking at Layer 24. The population probe (dashed orange) remains near zero throughout. Layer 28 shows a slight decline relative to Layer 24.
Figure 3: Cross-person transfer matrix (Qwen 2.5 7B, mean pupil size, Layer 24, PCA 50). Each cell (i,j)(i,j) shows Spearman ρ\rho when participant ii ’s probe predicts participant jj ’s mean pupil size. The diagonal (self ρ=0.369\rho=0.369 ) is consistently higher than the off-diagonal (other ρ=0.143\rho=0.143 ), confirming person-specificity. LLaMA 3.1 8B shows the same pattern (Table 4 ).
FeaturePerson ρ\rhoPerson ρ\rhoPop ρ\rhopp
 (raw)(confounds removed)(confounds removed)(confounds removed)
TRT_g20.1830.1690.0578.2×10−88.2\times 10^{-8}
TRT_g10.1510.1410.0653.5×10−63.5\times 10^{-6}
meanPupilSize0.2580.1810.1181.9×10−71.9\times 10^{-7}
TRT_b20.1530.1240.0683.1×10−53.1\times 10^{-5}
TRT_a10.1250.1070.0703.5×10−53.5\times 10^{-5}
Table 2: Person-specific probe performance before and after removing word-level confounds (word frequency, length, sentence position, GPT-2 surprisal) via per-participant nuisance regression. The person-specific advantage survives for all five features ( p<10−4p<10^{-4} ). Raw values come from re-running probes within the confound-control pipeline (same CV folds as the residualized condition) and may differ slightly from Table 1 .
ConditionPerson ρ\rhoPop ρ\rhoΔ\Deltapp
Full model (TRT_g2)0.1830.020+0.1633.1×10−93.1\times 10^{-9}
Shuffled (word ↔\leftrightarrow EEG scrambled)0.0010.001≈{\approx} 0.000†
nFixations (neg. ctrl.)0.2340.231+0.0040.360
Random proj. (50 random dirs.)0.1790.014+0.165‡
GloVe 50d (static, no LLM)0.057−- 0.006+0.0634.8×10−74.8\times 10^{-7}
Random emb. (no LLM)0.0100.004+0.006§
Table 3: Specificity controls for TRT_g2 at Layer 24, PCA 50. Shuffled : word-to-EEG pairings randomly scrambled (10 permutations). Random proj. : PCA’s top-50 directions replaced with 50 random orthogonal directions. GloVe 50d : LLM activations replaced with static GloVe word vectors ( Pennington et al., 2014 ) . Random emb. : LLM activations replaced with fixed random vectors per word. † Both collapse to zero; no meaningful comparison. ‡ Single run; paired test not applicable. § Both collapse to near zero; no meaningful comparison.
MetricQwen 2.5 7BLLaMA 3.1 8B
Best layer2424
TRT_g2 person ρ\rho0.1830.196
TRT_g2 pop ρ\rho0.0200.012
meanPupilSize person ρ\rho0.2610.278
Self transfer ρ\rho0.3690.384
Other transfer ρ\rho0.1430.154
Table 4: Side-by-side comparison of person-specific probe metrics for Qwen 2.5 7B and LLaMA 3.1 8B (both at Layer 24, PCA 50). All key metrics replicate quantitatively.
Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

MetricN=12N{=}12N=30N{=}30Change
TRT_g2 person ρ\rho0.1580.183+16%
meanPupilSize person ρ\rho0.2650.261stable
Self transfer ρ\rho0.2270.369+63%
Other transfer ρ\rho0.0110.143increased
Transfer pp -value3.2×10−73.2{\times}10^{-7}5.8×10−205.8{\times}10^{-20}1013×10^{13}{\times} stronger
Split-half cosine0.7520.824+10%
Appendix
Table 5: Comparison of characterization metrics at N=12N{=}12 (ZuCo 1.0 only) vs. N=30N{=}30 (combined). All metrics improve or remain stable with more participants.
Figure 4: Person-specific advantage ( Δρˉ=ρˉperson−ρˉpop\Delta\bar{\rho}=\bar{\rho}_{\text{person}}-\bar{\rho}_{\text{pop}} ) for 40 EEG features at Layer 24. Stars indicate significance: * p<0.001p<0.001 , ** p<10−4p<10^{-4} , *** p<10−6p<10^{-6} . Gamma bands carry the strongest individual signal regardless of reading window.
FeatureCohen’s ddPerson ρ\rho 95% CIPop ρ\rho 95% CIpp
TRT_g21.53[0.155,  0.209][0.155,\;0.209][−0.002,  0.040][-0.002,\;0.040]3.1×10−93.1{\times}10^{-9}
meanPupilSize1.99[0.225,  0.300][0.225,\;0.300][0.050,  0.093][0.050,\;0.093]9.3×10−129.3{\times}10^{-12}
TRT_g11.17[0.127,  0.187][0.127,\;0.187][0.020,  0.066][0.020,\;0.066]5.1×10−75.1{\times}10^{-7}
TRT_a11.18[0.103,  0.152][0.103,\;0.152][0.034,  0.075][0.034,\;0.075]4.8×10−74.8{\times}10^{-7}
TRT_b21.08[0.118,  0.174][0.118,\;0.174][0.050,  0.087][0.050,\;0.087]2.0×10−62.0{\times}10^{-6}
Appendix
Table 6: Effect sizes and 95% bootstrap confidence intervals ( B=10,000B=10{,}000 ) for the person-specific advantage at Layer 24, Qwen 2.5 7B. Cohen’s dd is computed from the paired person-vs-population ρ\rho difference across 30 participants. All five features show large effects ( d>1d>1 ); the population probe CI for TRT_g2 includes zero, confirming that the population signal is negligible for this feature.
More in this topic

EEG Decoding

Electroencephalography (EEG) decoding maps brain-signal patterns to mental states, stimuli, speech, or intended actions.

See all 198 papers

Explore similar work

CardsList

  1. Deep LLM layers expose passage-level context in EEG during silent reading

    Oct 16, 2026Ingo Marquardt, Anthilia Alchanat, Priyanka JainEEG DecodingEEG-to-Text Decoding

    nubrain

  2. A Shared Valence Axis Across Modern LLMs and Human EEG: The Saturation Regularity

    May 28, 2026Yousef A. Radwan, Xuhui Liu, Kilichbek Haydarov +2Cross-Modal Representation LearningNeural Decoding

    King Abdullah University of Science and Technology (KAUST)

  3. Encoding EEG Signals to Examine Human-Like Next-Word Prediction Behaviour in Language Models

    Jul 17, 2026Boi Mai Quach, Binh T. Nguyen, Cathal Gurrin +1LLM Evaluation

    School of Computing, ML-Labs, Department of Computer Science, School of Computing, ADAPT Centre, Dublin City University, Ireland · VNUHCM – University of Science · School of Computing, ADAPT Centre, Dublin City University, Ireland