analiza aiqbench
This paper is about interpretability and cognitive science, using LLMs as a testbed for how human language processing might work. Its input attribution method is the reusable contribution: it identifies which specific words drive brain alignment, and it finds that brain alignment and next-word prediction depend on largely different word subsets. Readers should treat the findings as evidence about model-brain correspondence under this method's assumptions, not as proof of shared computational mechanisms.
Fine-grained Analysis of Brain-LLM Alignment through Input Attribution
arXiv:2510.12355v2 Announce Type: replace Abstract: Understanding the alignment between large language models (LLMs) and human brain activity can reveal computational principles underlying language processing. We introduce a fine-grained input attribution method to identify the specific words most important for brain-LLM alignment, and leverage it to study a contentious research question about brain-LLM alignment: the relationship between brain alignment (BA) and next-word prediction (NWP). Our findings reveal that BA and NWP rely on largely distinct word subsets: NWP exhibits recency and primacy biases with a focus on syntax, while BA prioritizes semantic and discourse-level information with a more targeted recency effect. This work advances our understanding of how LLMs relate to human language processing and highlights differences in feature reliance between BA and NWP. Beyond this study, our attribution method can be broadly applied to explore the cognitive relevance of model predictions in diverse language processing tasks.
Abstract:Understanding the alignment between large language models (LLMs) and human brain activity can reveal computational principles underlying language processing. We introduce a fine-grained input attribution method to identify the specific words most important for brain-LLM alignment, and leverage it to study a contentious research question about brain-LLM alignment: the relationship between brain alignment (BA) and next-word prediction (NWP). Our findings reveal that BA and NWP rely on largely distinct word subsets: NWP exhibits recency and primacy biases with a focus on syntax, while BA prioritizes semantic and discourse-level information with a more targeted recency effect. This work advances our understanding of how LLMs relate to human language processing and highlights differences in feature reliance between BA and NWP. Beyond this study, our attribution method can be broadly applied to explore the cognitive relevance of model predictions in diverse language processing tasks.
Computation and Language (cs.CL)
arXiv:2510.12355 [cs.CL]
(or arXiv:2510.12355v2 [cs.CL] for this version)
https://doi.org/10.48550/arXiv.2510.12355 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 234 cuvinte
Subiecte conexe
How do neural networks regress cubic polynomials? Apparently, they use a trick invented in Milan 500 years ago
LessWrong · 22 iul. 2026
PIRAMID: Progress and Plans
LessWrong · 27 iul. 2026
NeuroAlign: Hierarchical Multimodal Fusion of Dynamic and Structural Neuroimaging for MCI Analysis
arXiv · 9 iun. 2026
MyoMechanix: Biomechanically-Grounded Compositional Skilled Activity Understanding and Coaching
arXiv · 27 aug. 2026