Research map / Large language models / Language, safety and interpretability

Divergences between Language Models and Human Brains

NeurIPS 2024 · Large language models › Language, safety and interpretability · approach: Linguistic encoding in LLM representations · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Uses MEG responses to narrative stories and an LLM-based hypothesis proposer to find where GPT-2 XL embeddings fail to predict brain activity, then shows fine-tuning on social/emotional and physical commonsense QA data improves brain alignment.

Problem

Language models predict human brain responses well, but it is unclear where and how LM and human language processing diverge; prior work focused on similarities.

Approach

Fit ridge-regression encoding models from LM embeddings to MEG, compute per-word prediction error, and use an LLM proposer/verifier (GPT-3 and FLAN-T5-XXL) to find textual hypotheses explaining high-error words; identified social/emotional intelligence and physical commonsense, then fine-tune GPT-2 XL on related QA datasets to improve alignment.

Key results

• Social/emotional and physical commonsense emerged as the top hypothesis themes for GPT-2 XL on the Harry Potter MEG data (top hypothesis validity 0.250 for emotional intensity).
• Human behavioral experiment (10 participants, 1,400 trials) showed a preference for divergent sentences under top hypotheses versus bottom hypotheses (chi-square p=0.024).
• Fine-tuning GPT-2 XL on Social IQa and PiQA improved MEG prediction over the base model in the language-processing window (about 50-300ms for social, 50-275ms for physical), with social-word MSE reductions significant at 200-275ms.

Measured effects reported in the paper

Surprising findings

• Fine-tuning on the physical dataset also gave a marginal, non-significant improvement on social words, which the authors attribute to social content embedded in PiQA.
• Llama-2 7B's top hypotheses focused on physical actions and lacked social/emotional themes, unlike GPT-2 XL.

Limitations

• Only two narrative-story datasets were used, limiting generality.
• Social/emotion annotation agreement was modest (Krippendorff's alpha 0.54).
• Fine-tuned models' improvements were not tied to a general LM-loss improvement, but the fine-tuned models' multiple-choice accuracy was modest (54.86% on Social IQa).

Experimental setup

Datasets: Harry Potter MEG dataset (Chapters 9-10) The Moth story-listening MEG dataset Social IQa PiQA

Benchmarks: Harry Potter MEG dataset The Moth MEG dataset

Metrics: Pearson correlation mean squared error hypothesis validity perplexity / language-modeling loss

Baselines: base GPT-2 XL Llama-2 7B GPT-3 proposer

Builds on: Zhang et al. proposer-verifier hypothesis generation (Zhong et al. 2022) Schrimpf et al. brain-score style LM-brain alignment Hasson et al. shared computational objective

Closest papers by meaning

• Scaling and context steer LLMs along the same computational path as the human brain (NeurIPS 2025 · similarity 0.90)
• fMRI predictors based on language models of increasing complexity recover brain left lateralization (NeurIPS 2024 · similarity 0.88)
• Improving Semantic Understanding in Speech Language Models via Brain-tuning (ICLR 2025 · similarity 0.88)
• The Grounding Gap: How LLMs Anchor the Meaning of Abstract Concepts Differently from Humans (NeurIPS 2026 · similarity 0.87)
• Hierarchical Frequency Tagging Probe (HFTP): A Unified Approach to Investigate Syntactic Structure Representations in Large Language Models and the Human Brain (NeurIPS 2025 · similarity 0.87)
• The Emergence of Abstract Thought in Large Language Models Beyond Any Language (NeurIPS 2025 · similarity 0.87)