大脑活动模式或有助于增强LLM的演绎推理能力

September 3, 2026 feature Brain activity patterns could help sharpen LLM deductive reasoning by Ingrid Fadelli, Phys.org Ingrid Fadelli Author Meet our staff & contributors Learn about our editorial standards edited by Gaby Clark, reviewed by Robert Egan Gaby Clark Scientific Editor Meet our editorial team Behind our editorial process Robert Egan Senior Editor Meet our editorial team Behind our editorial process Editors' notes This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility: fact-checked peer-reviewed publication trusted source proofread The GIST Add as preferred source Glowing neural activations emerge from the human brain and flow into an artificial neural network, steering its internal representations onto structured pathways for robust reasoning. The image visualizes a bridge between biological and artificial intelligence, where brain activation patterns guide machine inference from within. Credits: Zhouchen Lin, Kai Du, Mingqing Xiao. Large language models (LLMs), the artificial intelligence systems underpinning the functioning of ChatGPT, Gemini and other similar conversational agents, are now widely used worldwide. In addition to processing, interpreting and generating texts, some of these models can solve basic logical problems and answer some user questions with striking accuracy. While various past studies assessed the reasoning capabilities of some LLMs, how these models encode information to make logical predictions is poorly understood. In particular, the extent to which their activity patterns while processing reasoning problems resemble those in areas of the human brain linked with reasoning remains unclear. Researchers at Peking University and Tsinghua University recently carried out a study comparing brain activity patterns during reasoning with the numerical patterns through which LLMs process information while generating logical answers to questions. Their paper, published in Nature Machine Intelligence, suggests that language models and the human brain represent deductive reasoning in similar, but not identical, ways, while also showing that LLM representations could potentially be steered using brain imaging data. "The research grew out of two related lines of thinking," Dr. Mingqing Xiao, the first author of the paper, told Tech Xplore. "The first came from neuroscience: in the human brain, language and thought appear to be dissociable, that is, the networks that support linguistic processing are largely distinct from those involved in reasoning. "However, large language models learn primarily from language, yet they display cognitive abilities that appear increasingly human-like. That made us wonder whether these models, using language as a medium, have indeed acquired some aspects of human-like cognition that go beyond language itself." The second area of study that inspired this work is rooted in NeuroAI. This is an interdisciplinary field that merges neuroscientific concepts or data with AI models, with the goal of improving understanding of both the brain and AI systems. "Some recent studies have found correlational alignment between the internal representations of AI models and those of the brain," said Xiao. "This gave us a method for studying the similarity, while it also led us to a further question: could we move from correlation to guidance? In other words, could human neural signals be used not only to measure similarities with a model, but also to intervene in the model and improve its capabilities?" Comparing brain activity with LLM internal representations The first objective of the team's study was to determine whether the internal representations of LLMs are somewhat aligned with activity observed in specific brain regions known to support reasoning and logical problem-solving. The second was to explore the possibility of influencing an AI model's internal representations using brain signals or data, with the goal of improving its reasoning capabilities and robustness. "We used a public functional magnetic resonance imaging (fMRI) dataset containing fMRI recordings collected from human participants as they were solving deductive reasoning problems," explained Xiao. "We extracted neural signals from the relevant brain regions as our measure of brain representations underlying human higher-order cognition. We then presented the same linguistic stimuli to 10 open-source LLMs, ranging from 1.5 to 72 billion parameters, and extracted the activations of their internal units as the models' representations." Instead of only examining activity in brain regions known to support language processing, Xiao and colleagues also studied regions previously linked with human reasoning. They compared this activity with the internal representations of LLMs that were asked to solve the same reasoning tasks. "We used neural predictivity, a standard approach in NeuroAI research, to compare the model and brain representations," said Xiao. "In simple terms, we asked how well a model's internal representation could predict the brain's representation through a linear mapping. Higher neural predictivity indicates a stronger relationship between the two representation spaces." Subsequently, the researchers developed a brain-guided framework that could be used to steer the internal representations of LLMs using brain activity patterns. They tested their approach on 10 LLMs and found that it successfully improved their performance on tasks that required deductive reasoning. Promising findings and future research directions Overall, the results of this study suggest that LLMs and some regions of the human brain encode aspects of deductive reasoning tasks in partially similar ways. In addition, the team showed that brain activity recordings could potentially be used to improve the deductive reasoning of language models. "The most notable finding is that LLMs and the human brain are not only partially aligned during deductive reasoning, but that brain signals can also be used to guide the models' internal representations and improve their performance on the task," said Xiao. "These improvements showed a degree of generalization and, importantly, were complementary to the improvements obtained through conventional language-based supervision." The insights gathered by Xiao and colleagues could potentially help improve both theoretical models of human reasoning and the deductive reasoning capabilities of LLMs. In the future, the researchers' approach could inspire other computer scientists to try to guide the internal processing of AI systems using brain imaging data or other neuroscientific observations. "Our future studies will explore additional cognitive abilities and other neural recording modalities," added Xiao. "Language models may still fall short on some capabilities that are difficult to supervise explicitly through language, and representation guidance may offer something that language cannot. On the other hand, fMRI provides valuable spatial information but is a slow signal, and modalities with higher temporal resolution could let us analyze more finely." Written for you by our author Ingrid Fadelli, edited by Gaby Clark, and fact-checked and reviewed by Robert Egan—this article is the result of careful human work. We rely on readers like you to keep independent science journalism alive. If this reporting matters to you, please consider a donation (especially monthly). You'll get an ad-free account as a thank-you. Publication details Mingqing Xiao et al, Beyond representational alignment with brain-guided language models for robust reasoning, Nature Machine Intelligence (2026). DOI: 10.1038/s42256-026-01278-w. Journal information: Nature Machine Intelligence Key concepts Large language modelsAI alignment Who's behind this story? Ingrid Fadelli Freelance journalist with BSc Psychology and MA International Journalism. Covers AI, robotics, neuroscience, and astrophysics since 2018. Full profile → Gaby Clark MA in English, copy editor since 2021 with experience in higher education and health content. Dedicated to trustworthy science news. Full profile → Robert Egan Bachelor's in mathematical biology, Master's in creative writing. Well-traveled with unique perspectives on science and language. Full profile → © 2026 Science X Network Citation: Brain activity patterns could help sharpen LLM deductive reasoning (2026, September 3) retrieved 8 September 2026 from https://techxplore.com/news/2026-09-brain-patterns-sharpen-llm-deductive.html This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only. Mind readers: How large language models encode theory-of-mind Nov 11, 2025 LLMs are becoming more brain-like as they advance, researchers discover Dec 18, 2024 Key units in AI models mirror human brain's language system May 12, 2025 Like human brains, large language models reason about diverse data in a general way Feb 19, 2025 Improving AI large language models helps them better align with human brain activity May 27, 2024 The cost of thinking: Reasoning models share aspects of information processing with human brains Nov 20, 2025 Memristor chip breaks the capacity limit of brain-inspired associative memory 5 hours ago Invention may bring smart bionic eye into sight 4 hours ago Whole-body expansion and contraction make a robot seem more alive 6 hours ago A tiny crystal tilt may bring greener magnetic memory closer to everyday devices 6 hours ago Tracking energy flow reveals voltage pattern that slashes waste heat by over 60% 8 hours ago Editorial proposes limits on generative AI's role in scientific research 12 hours ago