What Shapes Brain Alignment of Large Language Models during Dialogue? Insights from Language Production and Comprehension

Yuko Nakagi; Yuya Chiba; Yasue Kishino

cs.CLOct 16, 2026

What Shapes Brain Alignment of Large Language Models during Dialogue? Insights from Language Production and Comprehension

Authors: Yuko Nakagi, Yuya Chiba, Yasue Kishino

Organizations: NTT, Inc.

Abstract

Brain alignment of large language models (LLMs) offers a way to investigate representational similarities between LLMs and the human brain during language processing. While previous studies have mainly focused on passive language comprehension, little is known about such alignment during dialogue, in which speakers engage in both language production and comprehension. In this study, using fMRI data collected during dialogue, we examine brain alignment during language production and comprehension across LLM layers and brain regions. We assess its association with the intrinsic dimension (ID) of LLM representations and the contributions of several interpretable linguistic and communicative features. Our results reveal distinct patterns of brain alignment between production and comprehension, as well as strong associations with ID in brain regions associated with each processing mode. Furthermore, we show that the contributions of linguistic and communicative features also vary across processing modes, LLM layers, and brain regions.

Figures & tables

Figure 1: Brain Alignment of LLMs during Dialogue. a Brain alignment for production and comprehension in one participant (sub-OSU01), using layer 13 of Llama-3.1, projected onto the inflated (lateral and medial views on both sides) and flattened cortical surfaces (occipital areas at the center) of both hemispheres. Red and blue indicate production and comprehension, respectively. Voxels significant in either production or comprehension are colored (all colored voxels $P<0.05$ , FDR corrected). b Mean brain alignment across the cerebral cortex (top left) and within each ROI (remaining panels) for production and comprehension across Llama-3.1 layers, averaged across participants. Color intensity indicates context length, and circles indicate the peak layer for each condition.

Figure 2: Relationship between LLM representational complexity and brain alignment during dialogue. a Layerwise ID of Llama-3.1 representations for production and comprehension, averaged across participants’ dialogue. Shaded regions indicate the standard error of the mean. Red and blue indicate production and comprehension, respectively, and color intensity indicates context length. b Spearman rank correlations between layerwise normalized ID and layerwise whole-cortex-averaged brain alignment for production and comprehension at each context length. Correlation coefficients are averaged across all participants and LLMs in Fisher $z$ space. c Spearman rank correlations between layerwise normalized ID and layerwise ROI-averaged brain alignment.

		L1	L13	L26	L32
Prod	BC	32.21	7.81	12.94	12.12
	L	85.89	41.57	49.15	45.95
	POS	71.57	28.73	38.06	36.75
	DA	68.42	25.82	34.39	32.85
Comp	BC	27.63	8.36	9.06	7.03
	L	78.60	33.53	43.71	38.40

Table 1: Mean Percentage Decrease in Brain Alignment after Removing Interpretable Features. Percentage decrease (

\%

) in whole-cortex-averaged brain alignment of Llama-3.1 after removing each of the four feature types. Results are averaged across participants and shown for layers 1, 13, 26, and 32 for both production (Prod) and comprehension (Comp).

Figure 3: Effects of Interpretable Features Related to Dialogue. Flatmaps : For Layers 1, 13, and 32 of Llama-3.1, the feature whose removal caused the largest decrease in brain alignment at each voxel is mapped onto the cerebral cortex of sub-OSU01. Only voxels with significant brain alignment ( $P<0.05$ , FDR corrected) of at least 0.1 in the full model are colored. Box plots : The unique contributions of each of the four features to brain alignment in each ROI at the final layer of Llama-3.1 for sub-OSU01. Boxes indicate the median and interquartile range, red lines indicate the mean, and points represent individual voxels. Gray indicates brain alignment in the full model, while yellow, green, purple, and blue indicate the unique contributions of BC, L, POS, and DA, respectively.

Appendix figures & tables22 assets

Supplementary material from the paper’s appendix.

Appendix

Dialogue Act	Definition
INFORMATION	Utterances that provide objective information or convey information obtained from others, including hearsay.
SELF-DISCLOSURE	Utterances in which the speaker expresses personal information, opinions, or feelings.
BACKCHANNEL	Utterances that indicate that the speaker is listening or encourage the interlocutor to continue speaking.
QUESTION	Utterances in which the speaker asks the interlocutor for information or a response.
SYMPATHY	Utterances that express agreement with the interlocutor.
NON-SYMPATHY	Utterances that reject the interlocutor’s opinion or express an opposing view.

Appendix

Table 2: Dialogue act tags and their definitions.

Figure A.1: Voxelwise Brain Alignment of Llama-3.1, Gemma-2 and GPT-OSS during Dialogue. .

Figure A.2: Brain Alignment of Gemma-2 and GPT-OSS during Dialogue. Mean brain alignment across the cerebral cortex (top left) and within each ROI (remaining panels) for production and comprehension across Gemma-2 (left) and GPT-OSS (right) layers, averaged across participants. Color intensity indicates context length, and circles indicate the peak layer for each condition.

Figure A.3: Brain Alignment of Llama-3.1 during Dialogue for Each Participant. Mean brain alignment for each participant across the cerebral cortex (top) and within each ROI (remaining panels) for production and comprehension across Llama-3.1 layers. Color intensity indicates context length, and circles indicate the peak layer for each condition.

Figure A.4: Brain Alignment of Gemma-2 during Dialogue for Each Participant. Mean brain alignment for each participant across the cerebral cortex (top) and within each ROI (remaining panels) for production and comprehension across Gemma-2 layers. Color intensity indicates context length, and circles indicate the peak layer for each condition.

Figure A.5: Brain Alignment of GPT-OSS during Dialogue for Each Participant. Mean brain alignment for each participant across the cerebral cortex (top) and within each ROI (remaining panels) for production and comprehension across GPT-OSS layers. Color intensity indicates context length, and circles indicate the peak layer for each condition.

Figure A.6: Relationship between LLM representational complexity and brain alignment during dialogue. Layerwise ID of Gemma-2 (left) and GPT-OSS (right) representations for production and comprehension, averaged across participants’ dialogue. Shaded regions indicate the standard error of the mean. Red and blue indicate production and comprehension, respectively, and color intensity indicates context length.

Models	Layers	Production								Comprehension
		BC		L		POS		DA		BC		L		POS		DA
		Before	After	Before	After	Before	After	Before	After	Before	After	Before	After	Before	After	Before	After
Llama-3.1	1	.38	-.00	.33	.01	.72	.01	.41	.01	.36	-.02	.32	.01	.70	.01	.41	.01
	13	.43	-.01	.30	.01	.75	.01	.51	.01	.41	-.03	.29	.01	.74	.02	.51	.00
	26	.34	-.00	.26	.01	.71	.03	.48	.00	.34	-.03	.26	.01	.71	.03	.48	.00
	32	.28	.01	.23	.01	.68	.05	.45	.02	.29	-.02	.23	.01	.68	.05	.46	.01

Appendix

Table 3: Comparison of ridge regression probing performance (Pearson’s

r

) before and after removing interpretable features across LLM layers for production and comprehension.

		L1	L12	L24	L42
Prod	BC	30.64	13.69	7.87	12.78
	L	86.34	66.38	48.52	55.75
	POS	66.97	45.44	28.87	38.76
	DA	67.35	40.80	25.72	35.56
Comp	BC	27.67	12.84	8.45	10.36
	L	78.60	56.47	41.19	45.90

Appendix

Table 4: Mean Percentage Decrease in Brain Alignment after Removing Interpretable Features. Percentage decrease (

\%

) in whole-cortex-averaged brain alignment of Gemma-2 after removing each of the four feature types. Results are averaged across participants and shown for layers 1, 12, 24, and 42 for both production (Prod) and comprehension (Comp).

		L1	L8	L17	L24
Prod	BC	33.38	15.72	9.87	15.61
	L	89.16	69.60	52.24	58.88
	POS	69.67	45.53	31.82	41.85
	DA	72.14	42.11	30.40	37.53
Comp	BC	29.17	13.77	9.68	9.87
	L	85.86	56.78	43.18	50.15

Appendix

Table 5: Mean Percentage Decrease in Brain Alignment after Removing Interpretable Features. Percentage decrease (

\%

) in whole-cortex-averaged brain alignment of GPT-OSS after removing each of the four feature types. Results are averaged across participants and shown for layers 1, 8, 17, and 24 for both production (Prod) and comprehension (Comp).

Figure A.7: Distribution of Dominant Interpretable Dialogue-Related Features for Brain Alignment during Production in Llama-3.1. For Layers 1, 13, and 32 of Llama-3.1, the feature whose removal caused the largest decrease in brain alignment at each voxel is mapped onto the cerebral cortex of each participant. Only voxels with significant brain alignment ( $P<0.05$ , FDR corrected) of at least 0.1 in the full model are colored.

Figure A.8: Distribution of Dominant Interpretable Dialogue-Related Features for Brain Alignment during Comprehension in Llama-3.1. For Layers 1, 13, and 32 of Llama-3.1, the feature whose removal caused the largest decrease in brain alignment at each voxel is mapped onto the cerebral cortex of each participant. Only voxels with significant brain alignment ( $P<0.05$ , FDR corrected) of at least 0.1 in the full model are colored.

Figure A.9: Unique Contributions of Interpretable Dialogue-Related Features to Brain Alignment of Layer 32 of Llama-3.1 during Production. Boxes indicate the median and interquartile range, red lines indicate the mean, and points represent individual voxels. Gray indicates brain alignment in the full model, while yellow, green, purple, and blue indicate the unique contributions of BC, L, POS, and DA, respectively. Asterisks mark significant differences between features ( $P<0.001$ , two-sided Wilcoxon signed-rank tests with Holm correction).

Figure A.10: Unique Contributions of Interpretable Dialogue-Related Features to Brain Alignment of Layer 32 of Llama-3.1 during Comprehension. Boxes indicate the median and interquartile range, red lines indicate the mean, and points represent individual voxels. Gray indicates brain alignment in the full model, while yellow, green, purple, and blue indicate the unique contributions of BC, L, POS, and DA, respectively. Asterisks mark significant differences between features ( $P<0.001$ , two-sided Wilcoxon signed-rank tests with Holm correction).

Figure A.11: Distribution of Dominant Interpretable Dialogue-Related Features for Brain Alignment during Production in Gemma-2. For Layers 1, 24, and 42 of Gemma-2, the feature whose removal caused the largest decrease in brain alignment at each voxel is mapped onto the cerebral cortex of each participant. Only voxels with significant brain alignment ( $P<0.05$ , FDR corrected) of at least 0.1 in the full model are colored.

Figure A.12: Distribution of Dominant Interpretable Dialogue-Related Features for Brain Alignment during Comprehension in Gemma-2. For Layers 1, 24, and 42 of Gemma-2, the feature whose removal caused the largest decrease in brain alignment at each voxel is mapped onto the cerebral cortex of each participant. Only voxels with significant brain alignment ( $P<0.05$ , FDR corrected) of at least 0.1 in the full model are colored.

Figure A.13: Unique Contributions of Interpretable Dialogue-Related Features to Brain Alignment of Layer 42 of Gemma-2 during Production. Boxes indicate the median and interquartile range, red lines indicate the mean, and points represent individual voxels. Gray indicates brain alignment in the full model, while yellow, green, purple, and blue indicate the unique contributions of BC, L, POS, and DA, respectively. Asterisks mark significant differences between features ( $P<0.001$ , two-sided Wilcoxon signed-rank tests with Holm correction).

Figure A.14: Unique Contributions of Interpretable Dialogue-Related Features to Brain Alignment of Layer 42 of Gemma-2 during Comprehension. Boxes indicate the median and interquartile range, red lines indicate the mean, and points represent individual voxels. Gray indicates brain alignment in the full model, while yellow, green, purple, and blue indicate the unique contributions of BC, L, POS, and DA, respectively. Asterisks mark significant differences between features ( $P<0.001$ , two-sided Wilcoxon signed-rank tests with Holm correction).

Figure A.15: Distribution of Dominant Interpretable Dialogue-Related Features for Brain Alignment during Production in GPT-OSS. For Layers 1, 17, and 24 of GPT-OSS, the feature whose removal caused the largest decrease in brain alignment at each voxel is mapped onto the cerebral cortex of each participant. Only voxels with significant brain alignment ( $P<0.05$ , FDR corrected) of at least 0.1 in the full model are colored.

Figure A.16: Distribution of Dominant Interpretable Dialogue-Related Features for Brain Alignment during Comprehension in GPT-OSS. For Layers 1, 17, and 24 of GPT-OSS, the feature whose removal caused the largest decrease in brain alignment at each voxel is mapped onto the cerebral cortex of each participant. Only voxels with significant brain alignment ( $P<0.05$ , FDR corrected) of at least 0.1 in the full model are colored.

Figure A.17: Unique Contributions of Interpretable Dialogue-Related Features to Brain Alignment of Layer 24 of GPT-OSS during Production. Boxes indicate the median and interquartile range, red lines indicate the mean, and points represent individual voxels. Gray indicates brain alignment in the full model, while yellow, green, purple, and blue indicate the unique contributions of BC, L, POS, and DA, respectively. Asterisks mark significant differences between features ( $P<0.001$ , two-sided Wilcoxon signed-rank tests with Holm correction).

Figure A.18: Unique Contributions of Interpretable Dialogue-Related Features to Brain Alignment of Layer 24 of GPT-OSS during Comprehension. Boxes indicate the median and interquartile range, red lines indicate the mean, and points represent individual voxels. Gray indicates brain alignment in the full model, while yellow, green, purple, and blue indicate the unique contributions of BC, L, POS, and DA, respectively. Asterisks mark significant differences between features ( $P<0.001$ , two-sided Wilcoxon signed-rank tests with Holm correction).

Representation Geometry in Language Models

Representation geometry in language models studies how embeddings and hidden states organize concepts, computation, and behavior.

See all 417 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
Fine-grained Analysis of Brain-LLM Alignment through Input Attribution
Oct 14, 2025Michela Proietti, Roberto Capobianco, Mariya TonevaLLM Interpretability Next-Token Prediction
Sapienza University of Rome, Italy · Sony AI, Zurich, Switzerland · Max Planck Institute for Software Systems, Germany
Brain-LLM Alignment Tracks Training Data, Not Typology
May 21, 2026Dongxin Guo, Jikun Wu, Siu Ming YiuMultilingual Language Models Cross-Lingual Representation Alignment
The University of Hong Kong Hong Kong, China · Stellaris AI Limited Hong Kong, China

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

Representation Geometry in Language Models

Explore similar work

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

Fine-grained Analysis of Brain-LLM Alignment through Input Attribution

Brain-LLM Alignment Tracks Training Data, Not Typology