Language models align with brain regions that represent concepts across modalities

Thank you very much for your feedback and suggestions!

Title: We agree that the title was not very clear, so we decided to change it to “Language models align with brain regions that represent concepts across modalities” (if that’s unclear as well, please let us know).

The previous title was trying to convey the main question of our brain encoding experiment (Fig. 4): are LM representations more aligned with voxels that represent language (language-selective ones) or with voxels that represent concepts (semantically consistent ones)?

Encoding performance by layer: These graphs show the layer-wise cross-validation scores per LM: with mean pooling over tokens, with last-token pooling. We have added it to the paper per your suggestion. For most models, it is consistent with the observations of Caucheteux & King (2022) and Tuckute et al. (2024), with middle layers performing the best. Interestingly, for some models, predictivity dips in the middle—we will investigate this in a follow-up analysis.

Additional analyses: Thank you for the suggestions! The cross-modal encoding would indeed be very interesting to try, and we will work on adding it to the paper. As for the GloVe/CLIP baselines, we already do something similar—like Caucheteux & King (2022), we look at the encoding performance at layer 0, which outputs the static word embeddings (and, for VLMs, the projected features from the ViT/CLIP vision encoder). The mean-pooled graph linked above shows the models’ predictivity at layer 0, and we will clarify this point in the paper.

The observed increase in alignment between VLM and ventral temporal, it does not directly demonstrate that the models capture cross-modal conceptual meaning.

To clarify, we never interpreted it this way—the reason we think LMs might capture cross-modal meaning is that they align better with voxels that are more semantically consistent (i.e., represent concepts across modalities).

Inter-subject consistency: As we understood, ISC is crucial for Bonnasse-Gahot & Pallier (2024) because they predict average activation across subjects in each voxel, while we predict activations in each subject individually and then aggregate (error bars/intervals in Fig. 3-5 are over subjects). The voxel populations compared in Fig. 4 (quartiles) are also different for each subject. In this context, we think that reliability of each voxel’s activation across subjects is less of a concern, but we will be happy to discuss further.

Thank you very much for your review and especially the pointers to the additional literature; we have now added these references!

Lacking discussion on what is known from neuroscience about the identified ROIs and whether the fact that they exhibit the highest semantic consistency is surprising or consistent with previous findings about the human brain.

We deliberately avoided drawing connections to prior studies that have implicated these or similar anatomical areas in semantic or other processes. Across-study comparisons based on macroanatomical areas—albeit common in some subfields of neuroscience—are precarious (e.g., Poldrack, 2011). In particular, macroanatomical areas are not ‘natural kinds’ and often contain two or more functionally distinct areas (see e.g., Fedorenko & Blank, 2020 for the example of “Broca’s area”), so simply because two studies find activation in a similar anatomical location (say, the “anterior IFG” or “posterior MTG”) does not imply that we are talking about the “same” functional area. Engaging in such speculation has led to a lot of misguided theorizing and flawed inferences in neuroscience. This said, the locations of our semantically-consistent areas are broadly consistent with putatively amodal semantic areas in Ivanova (2022, Ch. 5) and Popham et al. (2021). But determining whether these are indeed the same functional areas will require conducting the same experiments within the same individuals, so the activations can be compared directly.

Since I am not perfectly familiar with the related literature, I cannot judge whether the semantic consistency metric is really novel or not. What similar metrics have been introduced before, and why are they insufficient? This is something that should be covered in related work. It just seems rather obvious. I am not saying it is not novel enough or too simple, just that I would imagine others to have introduced similar metrics before.

To the best of our knowledge (including those of us with expertise in cognitive neuroscience of language), such a metric has not been introduced before (in large part, because, as far as we know, Expt 1 dataset from Pereira et al. is the only dataset that estimates responses to individual concepts across modalities).

Thank you very much for your feedback, we’re glad you’ve enjoyed our paper!

How is strong and weak response defined?

The strength of the voxel’s response is its $\beta$ value (L111) corresponding to a specific stimulus. We don’t formally quantify “strong” or “weak” in this paper, but in follow-up analyses (discussed below) we rank concepts by how much a voxel responds to them (mean $\beta$ over all stimuli for the concept) and consider the top $N$/bottom $N$ concepts to have a strong/weak response.

What is the interpretation of a voxel that consistently responds to one concept (e.g. bird), but not another (e.g. art)?

It is difficult to provide a definitive interpretation of what a specific voxel/neural population does. How meanings are represented in neural tissue remains one of the biggest mysteries in the field. One question is whether particular preferences can be explained by a lower-dimensional representation of the semantic space. Indeed, in some of our follow-up experiments, we seem to find that voxels fall into two broad categories: concrete-concept-preferring voxels and abstract-concept-preferring voxels (in line also with other recent work: Hoffman & Bair, 2025). There are likely other dimensions that underlie these preferences, to be discovered in future work.

Comparison to specific prior work

The data from Pereira et al. (2018) is indeed widely used, but most studies (including the mentioned Schrimpf et al., 2021; Sun et al., 2022; Oota et al., 2022; Aw et al., 2024) use Pereira’s Experiments 2-3 rather than Experiment 1. These are different datasets: Expt 1 probes concept understanding across contexts/modalities and Expts 2-3 focus on reading comprehension. Only Expt 1 has multimodal stimuli and yields data that can be used to identify semantically consistent voxels.

Thank you for the pointers to Oota et al. (2022a, 2022b), which use Expt 1 data; these are highly relevant, and we have now added these references. However, these studies ask distinct research questions from the one tackled in the current study; in particular, they don’t target voxels that represent concepts, focusing instead on brain networks with other functions (language, visual cortex, DMN, TPN).

Also, prior studies have already explored the use of multiple views to perform concept understanding in a context and whether the voxels are semantically consistent with language model representations across views [Oota et al. 2022b]

As we understand, Oota et al. (2022b) find voxels where you can train a concept word embedding decoder on one view (e.g., sentences) and test on another (e.g., pictures) with high pairwise accuracy. Compared to our semantic consistency metric, this is less direct evidence for whether these voxels represent meanings cross-modally. Also, in contrast to Oota et al., our study is an encoding and not decoding study: our encoding approach allows us to investigate how the feature spaces of different models are able to account for the semantically consistent voxels in the brain. Through our encoding approach, we can identify the most brain-like model, which can then serve as the foundation for exploring how it captures conceptual meaning across different modalities.

The use of multimodal models is required so we can embed the picture stimuli (Schrimpf et al.’s results with GPT-2 are for Pereira Expts 2-3, which are text-only). We use the newer and larger LMs in addition to GPT-2 because we assume those are of interest to the COLM community, and prior work (Aw et al., 2024) uses models of comparable sizes.

Thank you very much for the constructive criticism—we look forward to further discussion!

Research question

Unlike much existing work on LM–brain alignment, we use LM representations to predict responses specifically in the semantically consistent brain voxels (which we consider to be a reasonable proxy for the brain’s hypothetical “concept representation network”). We are not asking if the models and the brain represent language the same way, we are asking if they represent concepts in context similarly (and if LMs capture amodal conceptual meaning at all). This is our research question, and the novelty is in how we approach it empirically, identifying the “concept-responsive” voxels and measuring alignment in those voxels.

This distinction is also why we made the specific choices you highlighted:

The choice of brain regions of interest (ROIs): language vs. concepts in the brain

We introduce semantic consistency and use it to define ROIs because our focus is on brain areas that plausibly represent concepts. Language processing and semantic/conceptual processing are dissociated in the mind and brain. (1) Preverbal infants and patients with severe aphasia can make complex inferences about the physical and social worlds (e.g., Spelke, 2022; Dickey & Warren, 2015). (2) Semantic processing can be impaired (e.g., in patients with semantic dementia) without concurrent language deficits (for more discussion and references on this and the previous point, see this section in Reilly et al., 2025). (3) Brain imaging studies reveal distinct brain areas engaged by linguistic processing vs. abstract semantic processing (Ivanova et al., 2022; Wurm & Caramazza, 2019). However, debates continue as to where and how meanings are represented in the brain, so there are no “established” meaning-processing ROIs (cf. the language network). As a result, we leverage a unique dataset where a set of concepts are presented in three different contexts and develop a novel metric based on consistent preferences for particular concepts regardless of whether a concept is shown pictorially, in the context of related single words, or in a sentence context. We use the 3 brain areas that exhibit the highest semantic consistency as our ROIs. Two of these areas are distinct from the language network, showing little to no response to the language localizer; one shows some overlap with the language network and may serve as a gateway between the language system and the more abstract semantic areas. However, we agree that it is interesting to compare brain encoding performance in the language regions with that in our ROIs: we provide this plot (showing performance per Glasser area with >25% overlap with a left-hemisphere language region) for reference and are working on more follow-up analyses.

Informative vs. semantically consistent voxels

Informative voxels (as defined in Pereira et al., 2018) are ones whose responses to sentences/pictures/word clouds for a concept contribute to the ability to decode the embedding of the concept word. Semantically consistent voxels are ones that consistently show strong responses to some concepts and weak responses to others. Although these two measures may be correlated (at least when informative voxels are calculated from the average activations across paradigms, as in Pereira et al., 2018, but not in Oota et al., 2022b), they are not equivalent. Critically, our semantic consistency metric is a better proxy for how well a voxel represents meanings of concepts: (1) it is theoretically motivated from first principles of what it would mean for a neural population to represent a concept; (2) it relies on brain data only (like a localizer task) and is not tied to decoding or to any particular LM or embedding method; (3) we can study its relationship with LM predictivity (encoding quality) without making the analysis circular (as “informativeness” ≈ decoding quality).