A Low-Rank Space Captures Linear Alignment Between LLM and Human Neural Activity During Naturalistic Language Listening

Zach Paris; Emily Finn

cs.CLOct 29, 2026

A Low-Rank Space Captures Linear Alignment Between LLM and Human Neural Activity During Naturalistic Language Listening

Authors: Zach Paris, Emily Finn

Organizations: Department of Psychological and Brain Sciences, Dartmouth College, Hanover, NH 03755, USA

Abstract

Large language models (LLMs) are increasingly used as computational models of human language processing, yet it remains unclear which components of their representations drive alignment with neural activity. Here, we use ridge reduced-rank regression to identify the brain-predictive subspace of LLM representations: the representational directions that support linear prediction of neural responses. Across three naturalistic language-listening datasets spanning intracranial electrophysiology and fMRI, and six model conditions spanning autoregressive, bidirectional masked, and randomly initialized language models, we find that full linear encoding performance can be recovered using only 1-3% of the original embedding dimensionality. These low-dimensional subspaces generalize across held-out participants and, to a lesser extent, across datasets, suggesting that neural predictivity is concentrated in a compact and partially shared representational space. We next compare these brain-predictive directions with the principal axes of variation in the embeddings. Principal component analysis requires substantially more dimensions to recover neural prediction, demonstrating that the directions most useful for predicting neural activity are distinct from those capturing the most overall variance within LLM representations themselves. Finally, characterizing the learned subspaces allows us to surface interpretable features that are overrepresented in neural activity above and beyond the LLM representations. Together, these results show that LLM-brain alignment is supported by a small, structured subset of model representations and provides a framework for better understanding exactly why these representations align. Code can be found at: https://github.com/thefinnlab/low-rank-alignment

Figures & tables

Figure 1: (a) We view alignment as a measure of the linear features shared by the LLM embedding space and the neural response space. Our goal is to selectively extract the intersection of the two spaces. (b) Our pipeline. We learn two low-rank representations of the model embeddings and use them to predict neural data. The PCA space (green) prioritizes capturing the directions of principal variation in the LLM embeddings. The RRR space (blue) prioritizes the directions which best align between the two representational spaces teasing out the aligned subspace. (c) To identify the size of the aligned subspace we sweep the subspace rank, fit new models and re-evaluate. We find the point at which our RRR space performance saturates and compare it to the PCA point of saturation. (d) Leave-one-subject-out experiment. We fit the RRR space on all but one subject and then transfer the projection matrix to a held-out subject and fit a new mapping to that subject’s neural data. (e) We again sweep over ranks and compare the point of saturation on the held-out subject between the RRR space fit on all subjects, RRR space fit on all but the held-out subject, and the unsupervised PCA space.

Figure 2: Low-rank RRR recovers neural predictivity more efficiently than PCA. (a) Example rank-performance curves for GPT-J-6b layer 21 across the three datasets. (b) Mean convergence across model-layers showing RRR converges quicker than PCA. Generated by normalizing by max performance and enforcing monotonicity before averaging. (c) Saturation ranks for matched model–layer combinations, defined as the minimum rank at which all informative target-performance quantiles reach 99% of their maximum. RRR saturates at substantially lower rank than PCA across all three datasets. See Methods and Appendix for threshold and significance details.

Figure 3: Brain-predictive subspaces generalize across held-out participants. (a) Example rank–performance curves for within-participant RRR, transferred RRR learned from the remaining participants, and PCA. (b) Mean convergence curves across models and layers. (c) Saturation ranks for matched model–layer combinations. Transferred RRR reaches near-maximal performance at substantially lower rank than PCA and is generally comparable to within-participant RRR.

Figure 4: Brain-predictive subspaces partially generalize across datasets. (a) Mean convergence curves for within-dataset RRR, cross-dataset transferred RRR, and transferred PCA. (b) Summary of rank–performance efficiency across directed source–target dataset pairs. Transferred RRR is less efficient than within-dataset RRR but consistently more efficient than transferred PCA, indicating partial cross-dataset generalization.

Figure 5: Cross-dataset RRR transfer loss covaries with stimulus-dependent PCA shift. Across models, layers, and directed dataset pairs, the gap between within-dataset and transferred RRR performance is positively correlated with the corresponding PCA transfer gap ( $r=0.58$ , $n=184$ ). Each point represents a model–layer combination for one directed source–target transfer.

Figure 6: Linguistic feature enrichment differs across neural datasets. (a) Relative decoding advantage of RRR over PCA across ranks for example features. Positive values indicate greater decodability from RRR; negative values indicate greater decodability from PCA. Details of computation in A.3.1. (b) Feature-enrichment profiles are strongly correlated between the two iEEG datasets but not between either iEEG dataset and fMRI.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 7: We use all targets deemed significant by at least one model–layer combination. See table 1 for full counts.

Dataset	Model	0%	25%	50%	75%	100%
Podcast	random-gpt2	350	197	128	127	90
	glove-300	381
	roberta-large	463	626	538	450	515
	gpt2	275	435	467	535	541
	gpt2-xl-batch	337	415	526	565	516
	gpt-j-6b	450	630	638	574	540

Appendix

Table 1: Number of significant targets (electrodes for Podcast and Cars 2, voxels for fMRI) for each model-layer. Union represents the final number of targets used for our analysis by taking all targets which any model-layer deemed significant.

Figure 8: Mean encoding results of all of our model-layers across all of our datasets. We generally recover patterns typical of previous work where larger models outperform smaller models, and middle layers perform better than early or late layers. Performance is quite spread due to our liberal significance filtering.

Figure 9: GPT-J-6b layer 21 performance quantiles on different datasets. We can see that lower-quantiles generally lag a bit behind higher ones. We exclude any with a peak value less than 0.005 when calculating saturation to avoid unreliable targets leading to noisy overestimates.

Figure 10: Normalized performance curves for every quantile averaged across model-layers. All tested quantiles tend to converge quickly on RRR but less quickly for PCA.

Figure 11: Sensitivity analysis of how saturation rank changes as you change the required performance threshold across qualifying quantiles. Across all thresholds RRR converges quicker than PCA, but the effect becomes more drastic as the value increases. We selected 99% as our goal is to ask when max performance is almost entirely recovered.

Figure 12: Sensitivity analysis of how saturation rank changes as you change the floor correlation for a quantile to be included in the saturation rank computation. Using a floor of 0 causes some model layers to gain $\sim 1000$ saturation rank. Increasing to 0.005 leads to more reasonable results which we choose as this is so close to a noise response.

FEATURE	DATASET(S)	MEASUREMENT AND CALCULATION	CITATION(S)
Age of Acquisition	Podcast	Mean rated age at which the word is learned.	( Kuperman et al., 2012 )
Arousal	fMRI, Podcast	Rated emotional activation associated with the word.	( Mohammad, 2025 )
Association Entropy	Podcast	Entropy of the SWOW response-strength distribution for the cue word.	( De Deyne et al., 2019 ; Shannon, 1948 )
Concreteness	Cars 2, fMRI, Podcast	Mean rating of how tangible or perceptually accessible the word’s meaning is.	( Brysbaert et al., 2014 )
Content Word	Cars 2, fMRI, Podcast	Whether the word’s universal part-of-speech category is a content-word category.	( Nivre et al., 2020 ; Qi et al., 2020 )
Dependency Depth	Cars 2, fMRI, Podcast	Number of dependency-tree edges between the word and the sentence root.	( Nivre et al., 2020 ; Qi et al., 2020 )

Appendix

Table 2: Features decoded in the internal Cars 2, fMRI, and Podcast configurations.

Neural Encoding

Neural encoding studies how information is represented in biological neural activity and how models map stimuli or signals into neural codes.

See all 93 papers

Cross-lingual brain-language model alignment is robust but challenges hierarchical and computational accounts
May 20, 2026Ni Yang, Rui He, Philipp Homan +3Neural Encoding
Grammar and Cognition Lab, Department of Translation & Language Sciences, Universitat Pompeu Fabra, Barcelona, Spain. · Department of Adult Psychiatry and Psychotherapy, University of Zurich, Zurich, Switzerland. · Neuroscience Center Zurich, University of Zurich and ETH Zurich, Zurich, Switzerland. +3
Decomposing Brain-LLM Alignment with Interpretable Semantic Representations
Oct 15, 2026Camila Kolling, Mariya TonevaNeural Encoding LLM Interpretability
Max Planck Institute for Software Systems (MPI-SWS)
Heterogeneous Neural Predictivity from Language Models During Naturalistic Comprehension
Jun 25, 2026Xiao JiaNeuroimaging Neural Encoding
School of Artificial Intelligence The Chinese University of Hong Kong, Shenzhen

May 20, 2026J/K move · Enter open · S save

Cross-lingual brain-language model alignment is robust but challenges hierarchical and computational accounts

Ni Yang, Rui He, Philipp Homan, Iris Sommer+2

Grammar and Cognition Lab, Department of Translation & Language Sciences, Universitat Pompeu Fabra, Barcelona, Spain. · Department of Adult Psychiatry and Psychotherapy, University of Zurich, Zurich, Switzerland. · Neuroscience Center Zurich, University of Zurich and ETH Zurich, Zurich, Switzerland.+3

Neural Encoding

Open paper AskPDF arXiv

Brain-language model alignment is often interpreted as evidence that transformer models implement computations similar to those of the human brain. This assumes that neural predictivity reflects internal computational properties of large language models (LLMs), such as hierarchical contextual processing, predictive coding, or representational compression. An alternative possibility is that brain scores primarily reflect stable lexical-semantic correspondences shared by language models and the brain. Here we tested these interpretations using whole-brain encoding models across Mandarin, English, and French. Across all three languages, transformer representations significantly predicted activity in a distributed network spanning classical language regions, transmodal cortical systems, and subcortical structures. These spatial patterns showed substantial cross-linguistic overlap and remained remarkably stable across layers, providing little evidence that model depth systematically maps onto cortical processing hierarchies. Likewise, contextual transformer embeddings did not consistently outperform static lexical embeddings, despite providing some unique predictive variance. Finally, neither surprisal nor intrinsic dimensionality reproduced the layer-wise profile of brain scores, arguing against prediction and information compression as primary explanations for brain-LLM alignment. Together, these findings suggest that brain-LLM alignment is more robust across languages, transformer depth, and model architectures than previously appreciated, but less informative about shared computational mechanisms. Our results are more consistent with neural predictivity reflecting stable representational structure preserved across model transformations than with a one-to-one correspondence between their underlying computations.

Abstract

Neural Encoding

Explore similar work

Cross-lingual brain-language model alignment is robust but challenges hierarchical and computational accounts

Decomposing Brain-LLM Alignment with Interpretable Semantic Representations

Heterogeneous Neural Predictivity from Language Models During Naturalistic Comprehension