Research map / Large language models / Language, safety and interpretability
Hierarchical Frequency Tagging Probe (HFTP): A Unified Approach to Investigate Syntactic Structure Representations in Large Language Models and the Human Brain
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.
Introduces HFTP, a frequency-tagging probe that finds sentence- and phrase-selective MLP neurons in six LLMs and sEEG channels in 26 Chinese speakers, then aligns them via RSA to compare syntactic representations.
Problem
Unclear which computational units in LLMs encode hierarchical syntactic structure and whether they align with human brain mechanisms; prior methods make model-brain comparison of syntax methodologically inconsistent.
Approach
Adapts the hierarchical frequency tagging paradigm (1 Hz sentence, 2 Hz phrase rhythms) to LLM neuron activations and human sEEG via FFT and ITPC, with a word-order-randomized control, then uses SRDM-based RSA to align MLP neurons with brain channels.
Key results
- GPT-2 showed the highest left-hemisphere model-brain similarity (S = 0.654 for sentence corpus), with left-hemisphere alignment consistently exceeding right.
- Gemma 2 showed greater brain similarity than Gemma (S = 0.644 vs 0.582, L, sentence corpus), while Llama 3.1 showed lower alignment than Llama 2 (S = 0.514 vs 0.645, L, sentence corpus).
- Sentence-channel and phrase-channel proportions shift across brain ROIs: phrase channels decrease from A1 to IFG while sentence channels increase.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Upgrade from Gemma to Gemma 2 | Gemma family · sEEG-LLM SRDM alignment, sentence and phrase corpora | S(m,b) left hemisphere: 0.582 to 0.644 (sentence); 0.575 to 0.628 (phrase) | Helped |
| Upgrade from Llama 2 to Llama 3.1 | Llama family · sEEG-LLM SRDM alignment, sentence and phrase corpora | S(m,b) left hemisphere: 0.645 to 0.514 (sentence); 0.648 to 0.522 (phrase) | Hurt |
| Left-hemisphere alignment vs right-hemisphere alignment of LLMs | all six LLMs · sEEG-LLM SRDM alignment | S(m,b) | Helped |
| Word-order randomized control added to the corpus (isolating syntactic contribution) | LLM and sEEG analyses · HFTP peak detection | qualitative separation of experiment vs random spectra | No clear effect |
Surprising findings
- Llama 3.1 aligns less with the brain than its predecessor Llama 2, despite improved task performance.
- Sentence-channel and phrase-channel proportions across brain ROIs show no significant correlation, suggesting independent processing, unlike the LLMs where both levels share layers.
Limitations
- sEEG data collected primarily from Chinese speakers with Chinese stimuli; English evidence is preliminary from a single participant.
- Only a small set of model architectures and parameter scales was evaluated.
- Wikipedia and naturalistic results show attenuated peaks and weaker separation from controls.
Experimental setup
Datasets: Chinese and English syntactic corpora adapted from Ding et al. Chinese sentence and phrase corpora naturalistic text (dialogue, news, literature, poetry) Chinese and English Wikipedia 8-syllable/word corpora
Metrics: FFT amplitude permutation test 95% CI z-score Spearman correlation top-100 channel model-brain similarity S(m,b) ANOVA
Builds on: hierarchical frequency tagging (Ding et al.) representational similarity analysis knowledge neurons
Closest papers by meaning
- Brain-tuning Improves Generalizability and Efficiency of Brain Alignment in Speech Models (NeurIPS 2025 · similarity 0.88)
- A Polar coordinate system represents syntax in large language models (NeurIPS 2024 · similarity 0.88)
- Improving Semantic Understanding in Speech Language Models via Brain-tuning (ICLR 2025 · similarity 0.87)
- Divergences between Language Models and Human Brains (NeurIPS 2024 · similarity 0.87)
- Far from the Shallow: Brain-Predictive Reasoning Embedding through Residual Disentanglement (NeurIPS 2025 · similarity 0.87)
- TopoLM: brain-like spatio-functional organization in a topographic language model (ICLR 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.