Divergences between Language Models and Human Brains

Do machines and humans process language in similar ways? Recent research has hinted at the affirmative, showing that human neural activity can be effectively predicted using the internal representations of language models (LMs). Although such results are thought to reflect shared computational principles between LMs and human brains, there are also clear differences in how LMs and humans represent and use language. In this work, we systematically explore the divergences between human and machine language processing by examining the differences between LM representations and human brain responses to language as measured by Magnetoencephalography (MEG) across two datasets in which subjects read and listened to narrative stories. Using an LLM-based data-driven approach, we identify two domains that LMs do not capture well: social/emotional intelligence and physical commonsense. We validate these findings with human behavioral experiments and hypothesize that the gap is due to insufficient representations of social/emotional and physical knowledge in LMs. Our results show that fine-tuning LMs on these domains can improve their alignment with human brain responses.1

Paper

References (90)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer fRry7/10 · confidence 4/52024-06-20

Summary

The paper studies the differences in language representations in LLMs (GPT-2 XL) and the human brain (MEG activity), by using LLM representations to predict MEG activity. They use an LLM-based approach to identify two domains where LLM representations do not capture brain activity well: social/emotional intelligence and physical commonsense. They also show that fine-tuning GPT-2 XL on these domains can improve their alignment with human brain responses.

Strengths

From most to least significant: 1. The paper studies the highly interesting, important, and relevant topics of social/emotional intelligence and physical commonsense. Their results highlight the importance of improving social/emotional intelligence and physical commonsense to build more human-brain-aligned LLMs. 2. Interesting methodology using proposer and verifier LLMs. LLMs are growing increasingly capable, and it is great that the paper takes advantage of this trend, to advance the study of the brain. 3. After their analyses on brain alignment, I appreciate that they performed further investigations: human behavioral experiments (Section 3.3), and releasing annotations as a further resource to the dataset (Section 5).

Weaknesses

From most to least significant: 1. They used a relatively low layer (layer 7) of GPT2-XL when making the claim that LMs have poor brain alignment for high-level properties (social/emotional intelligence, physical commonsense). They should have studied a higher layer instead/too. - The paper argues they identified two domains that LMs do not capture well: social/emotional intelligence and physical commonsense. - However, they used layer 7 of GPT-2 XL, a relatively low layer since GPT-2 XL has 48 layers. They say they use this because it is the best at predicting brain activity, but this seems much lower than prior research: Layer 8 of 12 in BERT and Layer 12 of 19 in T-XL [1] - Prior work has shown that lower layers of LLMs capture lower-level linguistic properties, while higher layers capture higher-level properties. Prior work also showed that "LM representations can predict the hierarchy of brain responses" (lines 290-291 in their paper). - In this case, is it possible that the higher layers (30+) of the same GPT-2 XL model will achieve high brain alignment for social/emotional intelligence and physical commonsense? - However, re-evaluating more layers is expensive, and I am not requesting them to do this if they can address the comment in writing. 2. Results in Table 1 are not convincing in supporting their overall claims. The LLM-based system of proposers and verifiers generated hypotheses on how the top 100 least-predicted sentences differ from the top 100 well-predicted sentences. However, it seems that 331 of 708 human participants did not agree with the top hypotheses suggested by the LLM system (Table 1), i.e., either found the sentences to be "Equal" or "Convergent". This casts doubt that the two sets of sentences actually differ in terms of social/emotional intelligence and physical commonsense. This, in turn, casts doubt on their overall claim that LLMs do not capture social/emotional intelligence and physical commonsense well. 3. Additional training data as a potential confound - The paper showed that training LLMs on a dataset for emotional intelligence improves brain alignment. I agree that by itself this is interesting. However, is this because of: (A) statistical similarities between the text in this training dataset and the MEG stimuli, especially for words related to emotional intelligence, or (B) LLM gains better emotional intelligence understanding, as labels were provided and it learns to predict emotions correctly? - Alternative reasons that finetuning on social dataset improved brain alignment on social words: (a) because the model was provided additional training data? (b) because the additional dataset was higher quality? - Possible control experiment: training the model (using language modeling) on the same dataset but without associating each question with the correct label. Basically, do not tell the model which is the correct answer (no supervision). - It is possible that the authors can answer these questions using their further experiments in Appendix I and M. Still, I hope they explicitly address, within the paper, the issue of additional training data as a potential confound. [1] M. Toneva and L. Wehbe, “Interpreting and improving natural-language processing (in machines) with natural language-processing (in the brain).” 2019

Questions

1. Why is layer 7 of GPT-2 XL the best at predicting brain responses for this dataset? Layer 7 of 48 seems relatively low as opposed to intermediate, whereas prior research showed that intermediate layers are best (lines 115-116 in your paper).

Rating

7

Confidence

4

Soundness

2

Presentation

3

Contribution

3

Limitations

The limitations I identified have been mentioned in the paper.

Reviewer b2dQ7/10 · confidence 4/52024-07-12

Summary

Language models are known to predict MEG signals in humans during reading. In this work, the authors explored for what "kinds" of texts are language models bad at predicting MEG signals. The authors used a novel method to propose possible hypotheses, and found multiple possible categorizations of weak prediction texts. Focusing on social knowledge and spatial commonsense, the authors find that finetuning the language model on texts from these respective domains improves the prediction accuracy.

Strengths

1. The experiment design is clean and straightforward. 2. The detailed evaluation of error pattern during brain signal prediction is very valuable. 3. The automated method of hypothesis proposal might be generalizable to other forms of experiments in cognitive sciences as well.

Weaknesses

Many aspects mentioned in the work where the language model are not performant on, for example common sense and social reasoning, have large leaps forward in more modern models. A future work could focus on using more modern language models with stronger capacity.

Questions

1. Temporal signal prediction is convincing, another aspect that is interesting is the spectral structure of the predictions. 2. Do layers other than 7 display the same pattern of lower prediction accuracy? Do you observe this consistently across layers, or just at layer 7?

Rating

7

Confidence

4

Soundness

4

Presentation

4

Contribution

3

Limitations

Properly addressed.

Reviewer jRnr6/10 · confidence 4/52024-07-14

Summary

This paper explores the differences between LMs and the human brain in processing language. The authors conduct experiments using MEG data from reading and listening tasks to investigate elements of MEG responses that LMs cannot adequately explain. LLMs are used to automatically generate hypotheses, identifying the domains where LMs lack knowledge compared to the human brain. The authors then fine-tune LMs on these specific domains to improve the LMs-brain alignment.

Strengths

The idea of assessing and interpreting the biological plausibility of LMs is a feasible way to enhance both their interpretability and model design. Focusing on the divergences between LMs and the brain is currently lacking in the field. The paper is well-written, includes good visualizations, and the idea is easy to follow with good reproducibility.

Weaknesses

Only two LMs are used in the experiment. Additionally, GPT-2 is somewhat outdated compared to recently released open-source LLMs, as it shows limited language understanding and reasoning capability, making it far from achieving human-level AGI. The focus should be more on LLMs such as Mixtral and Gema. More details about the prompts and API usage should be discussed for the proposer and verifier LLMs.

Questions

The relatively narrow datasets used in the paper, focusing on human dialogue and stories, which intuitively consist of emotion representation, could easily introduce bias.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

See Weakness and Questions

Reviewer hoiu7/10 · confidence 4/52024-07-15

Summary

The authors use a data-driven method to generate hypotheses about particular words/linguistic contexts in which a brain encoding model does not accurately predict the brain response to language. they find that social/emotional/physical complexity, along with linguistic complexity, are all associated with worse brain encoding performance. They fine-tune the base GPT model to perform better at social/emotional reasoning tasks and find that this yields an improvement in brain encoding for these particular words within the relevant language processing window (~75-400 ms). Update: I have read the other reviews and author rebuttals and will keep my score as-is.

Strengths

- This is an interesting and novel approach using LLM proposer-verifier along with a behavioral experiment to validate the labeling generated by models. - The thorough experimental pipeline shows that a fine-tuning intervention designed along these hypotheses improves brain encoding performance as expected. - The paper opens up a new mysterious result to be explored -- it looks like there's a tradeoff between fitting these temporally intermediate responses and very early and very late responses to words.

Weaknesses

- The fine-tuning intervention needs to be appropriately baselined, for example by fine-tuning on other tasks which don't match the hypotheses about what is driving brain encoding performance. If fine-tuning on an appropriate counterfactual sample doesn't improve brain encoding performance, this will strengthen the positive finding. (I'm not totally clear on the discussion in Appendix M -- are you saying that fine-tuning through language modeling on Harry Potter doesn't yield brain encoding improvements? This is a step in the right direction, I think.) - The effect sizes of improvements here are very small; the results would be much stronger if you could reproduce this effect on a second dataset.

Questions

It seems like there is a lot more going on in the example sentences (Appendix C) than the hypotheses cover. For example, many function words are colored as "most divergent" despite not fitting the leading hypotheses. Can you provide a broader sample that might convince the reader of the coverage of the model-derived hypotheses?

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Reviewer fRry2024-08-08

I will raise my score from 6 to 7. The authors provided convincing clarifications for most of the weaknesses (and their sub-points) I raised. 1. Best Performing Layer (resolved) - Thanks for the clarification on MEG vs fMRI, and larger vs smaller models. 2. Human Study (resolved) - Thanks for the clarification, especially that "There are a total of ten hypotheses, and each sentence may only satisfy one or two of them." This helped me to understand why so many participant ratings are "Equal" or "Convergent", even for the set of top hypotheses. 3. Fine-Tuning (resolved) - Thanks for the clarifications.

Reviewer jRnr2024-08-12

I have read the authors' rebuttal which has addressed my concerns to some degree. I have raised my rating accordingly. Thanks.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC