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

The core experimental finding—that person-specific linear probes predict EEG power significantly better than population probes—is robust across multiple validations. The effect replicates in an independent LLaMA 3.1 8B model with nearly identical effect sizes ($\rho=0.196$ vs. $0.012$ for high-gamma at Layer 24), survives rigorous confound controls (word frequency, length, surprisal), and passes targeted negative controls including shuffled labels and fixation counts. The demonstration that these directions are non-transferable across individuals and temporally stable strongly suggests the probes capture genuine neural signatures. The authors note: "Self-prediction averages $\rho=0.369$, while cross-person prediction drops to $\rho=0.143$ ($p=5.8\times10^{-20}$, paired $t$-test)" and "The cosine similarity between the two weight vectors [split-half] averages $0.824$ ($p=6.6\times10^{-31}$)."

The absolute correlation values remain modest ($\rho \approx 0.18$), raising questions about practical utility for real-time brain-computer interfaces despite the large relative improvement over baselines. The study's sample size of 30 participants, while standard for simultaneous EEG and eye-tracking, limits generalization to broader populations. Additionally, the causal interpretation remains unestablished: the paper identifies correlational directions but does not demonstrate that manipulating these directions during generation would produce meaningful personalization, nor does it disentangle whether the signals reflect stable neuroanatomy or idiosyncratic reading behaviors. The authors acknowledge these limitations explicitly: "Our probes are linear, and nonlinear methods might capture additional structure, though they risk overfitting at the current data scale (roughly 5,000 words per participant). We do not attempt causal interventions: the person-specific directions are identified, not injected into the residual stream. The ZuCo dataset provides 30 participants, which is standard for simultaneous EEG and eye-tracking but modest for claims about the generality of individual differences."

The evidence supports the central claim that LLM activations contain person-specific neural information inaccessible to population-level analysis, with the fixation count negative control ($p=0.360$) providing a compelling baseline comparison. However, the comparison to static embeddings reveals nuance: while contextualized LLM representations show a person-specific advantage of $\Delta=0.163$, static GloVe embeddings show a smaller but significant advantage ($\Delta=0.063$), suggesting that individual differences are partially accessible even from non-contextual word vectors. The authors fairly contextualize their work against population-averaging paradigms like Brain-Score (Schrimpf et al., 2021) and representation engineering (Zou et al., 2023), accurately noting that prior work evaluates alignment at the population level or targets text-defined concepts rather than individual readers. As stated: "The GloVe control... uses static 50-dimensional word embeddings: a small but significant advantage emerges ($\Delta=0.063$, $p=4.8\times10^{-7}$), yet it is $2.6\times$ smaller than the LLM's ($\Delta=0.163$)."

The experimental pipeline is described with sufficient detail for reproduction, including specific model identifiers (Qwen/Qwen2.5-7B-Instruct, meta-llama/Llama-3.1-8B-Instruct), PCA dimensionality reduction to 50 components, Ridge regularization grid ($\alpha \in \{0.01,0.1,1,10,100,1000\}$), and 5-fold sentence-stratified cross-validation. The ZuCo 1.0 and 2.0 datasets are publicly available, and compute requirements (NVIDIA L4 GPU, 11-17 hours) and software versions (PyTorch 2.7.1, cuML 25.12) are documented. However, no code repository URL or persistent identifier is provided in the text, and the PCA projection is fit on the full vocabulary of 12,200 words, requiring exact reproduction of the tokenization pipeline. Without released preprocessing scripts or word-level activation extractions, independent reproduction of the exact feature representations may be challenging. The authors state: "PCA is fit on the full set of 12,200 unique words per layer, projecting to 50 dimensions. Ridge regression $\alpha$ is selected from $\{0.01,0.1,1,10,100,1000\}$ via held-out validation... 5-fold cross-validation is sentence-stratified."