Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Sparse Computational Components Contribute Disproportionately to Hierarchical Processing in Language Models
Poster Session F, Friday, October 2, 2:45 - 4:45 pm, Wangari Maathai
This poster is part of the Sandbox Series.
Shuguang Yang1,2,3, Feizhen Cao1,2,3, Junyi Li1,2,3, Ziyi Wang1,2,3, Yujing Nie1,2,3, Suiping Wang2; 1School of Psychology, South China Normal University, Guangzhou, China, 2Philosophy and Social Science Laboratory of Reading and Development in Children and Adolescents (South China Normal University), Ministry of Education, Guangzhou, China, 3Guangdong Key Laboratory of Mental Health and Cognitive Science, South China Normal University, Guangzhou, China
How structured computation is organized within distributed systems remains a central question in artificial intelligence and cognitive neuroscience. Hierarchical language processing offers a useful test case because it requires a system to combine discrete elements into structured internal representations while limiting representational load. Large language models (LLMs) show sensitivity to hierarchical linguistic structure, yet it remains unclear whether such computation is broadly distributed across the network or relies unevenly on a small subset of computational units. We tested this question at the level of individual units in multiple LLMs. We first constructed a Chinese natural-text corpus with controlled hierarchical organization and applied a hierarchical frequency-tagging probe to Chinese-pretrained Llama 2 and Llama 3. Candidate units were identified by their responses at phrase- and sentence-level structural frequencies, and a subsequent linear regression analysis tested whether their activations tracked syntactic depth after controlling for token length. We then examined whether the incremental representations carried by these units aligned with human brain activity measured with fMRI. Functional MRI data were collected from 35 Chinese participants while they read four-character phrases, and representational similarity analysis compared model and neural dissimilarity structures during incremental composition. Finally, we tested functional relevance through selective ablation and activation scaling on a Chinese reading-comprehension benchmark, with layer-matched random perturbations as controls. Hierarchical-structure sensitivity was concentrated in an extremely sparse subset of units. Only 1,337 units in Llama 2 and 1,681 units in Llama 3 met the selection criteria, corresponding to 0.39% and 0.37% of all units, respectively. These units carried linearly decodable information about syntactic depth beyond sequence length and were more densely distributed in deeper layers. Their incremental representations during phrase composition showed significant alignment with human brain activity across language-related regions, including inferior frontal and posterior temporal areas. Selective ablation of these units caused a marked decline in reading-comprehension performance, driving accuracy toward chance, whereas random ablation of matched units did not significantly impair performance. Activation scaling further modulated model behavior in a factor-dependent manner and improved Llama 2 performance within a narrow effective range; similar ablation effects and non-monotonic scaling patterns were observed in additional models, including Qwen and DeepSeek. These findings suggest that distributed representations do not necessarily imply uniformly distributed computational contributions. Hierarchical processing in LLMs can be embedded within broadly distributed representations while relying unevenly on a small set of identifiable computational components whose activity can be experimentally manipulated. By combining unit-level identification, model-brain alignment, and targeted intervention, this study provides an operational framework for examining how structured computation is organized within distributed artificial systems and offers a basis for comparing principles of functional organization across artificial and biological language systems.
Topic Areas: Computational Approaches,