Far from the Shallow: Brain-Predictive Reasoning Embedding through Residual Disentanglement

TL;DR

This work introduces a residual disentanglement framework to isolate higher-order reasoning representations in large language models and map them to human brain activity. By locating feature-specific LM layers and regressing out lower-level signals, the authors produce four orthogonal embeddings for lexicon, syntax, meaning, and reasoning, enabling targeted brain encoding with ECoG data. The findings show reasoning has a distinct, later neural signature and engages broader cortical regions beyond traditional language areas, while standard unsegmented embeddings are biased toward shallow features. Overall, the method reveals a hierarchical brain-LLM alignment and clarifies the neural substrates of linguistic reasoning beyond lexical processing.

Abstract

Understanding how the human brain progresses from processing simple linguistic inputs to performing high-level reasoning is a fundamental challenge in neuroscience. While modern large language models (LLMs) are increasingly used to model neural responses to language, their internal representations are highly "entangled," mixing information about lexicon, syntax, meaning, and reasoning. This entanglement biases conventional brain encoding analyses toward linguistically shallow features (e.g., lexicon and syntax), making it difficult to isolate the neural substrates of cognitively deeper processes. Here, we introduce a residual disentanglement method that computationally isolates these components. By first probing an LM to identify feature-specific layers, our method iteratively regresses out lower-level representations to produce four nearly orthogonal embeddings for lexicon, syntax, meaning, and, critically, reasoning. We used these disentangled embeddings to model intracranial (ECoG) brain recordings from neurosurgical patients listening to natural speech. We show that: 1) This isolated reasoning embedding exhibits unique predictive power, accounting for variance in neural activity not explained by other linguistic features and even extending to the recruitment of visual regions beyond classical language areas. 2) The neural signature for reasoning is temporally distinct, peaking later (~350-400ms) than signals related to lexicon, syntax, and meaning, consistent with its position atop a processing hierarchy. 3) Standard, non-disentangled LLM embeddings can be misleading, as their predictive success is primarily attributable to linguistically shallow features, masking the more subtle contributions of deeper cognitive processing.

Figures (9)

• Figure 1: a) Hierarchical representations in an LLM. Transformer layers accumulate information in order: lexical features emerge first, followed by syntax, contextual meaning and eventually higher-order reasoning, with still-richer knowledge continuing in later layers. b) Minimal-pair probing tasks. Three diagnostic sentence sets separately test syntax, concept meaning and multi-premise reasoning. c) Layer localization from probing curves. We define $L_s$ – the earliest layer where syntax performance saturates while meaning is still low; $L_m$ – layer where meaning saturates but reasoning has not yet emerged; and $L_r$ – layer where reasoning performance plateaus. These identified layers through probing will be used in later analyses. d) Feature disentangling across layers. Starting from the localized layers, we iteratively regress lower-level features out of higher ones. Details of residual embedding constructions could be found in Algorithm \ref{['algo:residual']}. Residual disentangling yields four orthogonal embeddings that isolate lexicon, syntax, meaning and reasoning information. e) Brain encoding with purified features. Each residual feature is fed into a ridge encoder to predict high-gamma ECoG responses of Podcast listening. Comparing predicted and actual neural signals reveals the spatiotemporal distribution of cortical activity uniquely associated with lexicon, syntax, meaning and reasoning representations.
• Figure 2: a) Pairwise cosine similarity among representations before (left) and after (right) residual disentanglement. The hidden states at feature saturation layers ($H_l, H_s, H_m, H_r$) exhibit substantial overlap. In contrast, the residual embeddings ($E_l, E_s, E_m, E_r$) show near-zero off-diagonal similarity. b) Feature probing results on hidden states (left) and residual embeddings (right), shown as accuracy. Each residual embedding achieves the highest performance on its corresponding task, while performing worse on unrelated tasks.
• Figure 3: a) Peak correlation across linguistic feature models. As described in the Methods section, electrodes with z > 3.95 (one-tailed $\alpha = .05$, Bonferroni-corrected across N = 1268 electrodes) were deemed responsive, corresponding to values exceeding 3.95 standard deviations above the shuffle mean. Boxplots indicate the median and interquartile range across activated electrodes with red lines indicating means additionally. Asterisks mark significant differences between adjacent features (Welch’s t-test; $p<0.001$), showing that lexical and syntax features yield significantly higher correlations than high-level linguistic representations. b) Peak correlation by subject and feature. The number of responsive electrodes for each subject is shown below the x-axis. Lexical features generally exhibit higher correlations, indicating stronger neural alignment for lower-level linguistic representations.
• Figure 4: a) Average correlation time courses for the top 10% most responsive electrodes reveal distinct temporal dynamics across linguistic features. Responsive electrodes were identified following the strategy described in the Methods section. Syntactic signals rise before word onset, followed by meaning, lexicon and reasoning, which peak latest at $\sim$362 ms. Significance markers (×) indicate time points where correlations are significantly greater than zero (one-tailed $t$-test, FDR-corrected $q<0.05$). b) Temporal profile of individual electrodes selective to each feature.: Lexicon, Syntax, Meaning, Reasoning, and Full. Electrodes were selected the same way as in panel a.
• Figure 5: a) Spatial distribution of feature-selective responsive electrodes (MNI space). Displayed electrodes are those that met the responsiveness criterion defined in the Methods section. For visualization, z-scores were clipped to 3–6 to prevent extreme values from dominating the color scale and obscuring spatial patterns. b) For each electrode, the feature with the highest peak z-score among Lexicon, Syntax, Meaning, and Reasoning models is shown as its dominant feature. Colored dots indicate the dominant model. Across all electrodes, the counts of dominant features were: Syntax = 166, Meaning = 161, Reasoning = 128, and Lexicon = 42.