Large Language Models Show Signs of Alignment with Human Neurocognition During Abstract Reasoning

Human Neurocognition During Abstract Reasoning: EEG and Eye-Tracking DatasetStudy OverviewThis dataset contains electroencephalography (EEG) recording, eye-tracking recording, and behavioral data from human participants performing an abstract reasoning task (visual analogy problem-solving). The original study investigated the alignment between Large Language Models (LLMs) and the human neural responses during abstract pattern completion and reasoning.Primary Research QuestionTo what extent do LLM representations reflect the neural mechanisms underlying human abstract reasoning?Reference PublicationsMain preprint:Pinier, C., Vargas, S. A., Steeghs-Turchina, M., Matzke, D., Stevenson, C. E., & Nunez, M. D. (2025). Large Language Models Show Signs of Alignment with Human Neurocognition During Abstract Reasoning. arXiv preprint arXiv:2508.10057.Conference publications:Pinier, C., Vargas, S. A., Steeghs-Turchina, M., Matzke, D., Stevenson, C. E., & Nunez, M. (2026, March 6). Large language models show signs of alignment with human neurocognition during abstract reasoning [Poster session]. ICLR 2026 Workshop - From Human Cognition to AI Reasoning: Models, Methods, and Applications. PDFPinier, C., Stevenson, C. E., & Nunez, M. D. (2025). Moderate evidence for large language models reflecting human neurocognition during abstract reasoning [Poster session]. Cognitive Computational Neuroscience (CCN) 2025, Amsterdam, Netherlands.Dataset DescriptionParticipantsNumber of participants: 25Recruitment: University of Amsterdam communityInclusion criteria: Native or fluent English speakers, normal or corrected-to-normal visionSessions: Multiple sessions per participant (1-5 sessions)Experimental TaskParticipants completed an abstract visual reasoning task involving pattern completion and analogy problems.Task Structure"Encoding phase" w/ individual icons of both pattern and response options flashing individually: 600 ms"Decision phase" w/ pattern and response options displayed w/ maximum response time: 12 secondsNumber of sequences per session: 80StimuliPattern types: 8 different abstract visual patterns representing varying levels of relational complexity:AAABAAAB, ABABCDCD, ABBAABBA, ABBACDDC, ABBCABBC, ABCAABCA, ABCDDCBA, ABCDEEDCDisplay: During the "Decision phase", the final icon in the sequence was replaced by a question mark. Four response icons were also displayed.Data CollectionEEG RecordingElectrode montage: BioSemi 64-channel standard montageSampling rate: 2048 HzAdditional channels:EOG (4 channels): EOGL, EOGR, EOGT, EOGBStimulus trigger channel: StatusRecording type: ContinuousPower line frequency: 50 HzEye-Tracking RecordingSampling rate: 2000 HzDevice: EyeLink 1000 PlusBehavioral DataRaw behavioral responses stored in TSV format including trial number, stimulus pattern ID, participant response, accuracy, and response time.Electrode PositionsNote on electrode coordinates: The standard BioSemi 64-channel montage does not include recorded electrode positions, as this study used a standard template montage (not subject-specific digitization). Electrode locations follow the standard 10-20 positioning system for BioSemi 64-channel caps.Ethics and Study ApprovalThis research project complies with the guidelines formulated by the Ethics Review Board (FMG-UvA), University of Amsterdam, The Netherlands, and has been approved by the aforementioned Ethics Review Board on 19-06-2024.LicenseThis dataset is made available under the Creative Commons Attribution 4.0 License (CC BY 4.0).How to Cite This DatasetAssociated Publication: Pinier, C., Vargas, S. A., Steeghs-Turchina, M., Matzke, D., Stevenson, C. E., & Nunez, M. D. (2025). Large Language Models Show Signs of Alignment with Human Neurocognition During Abstract Reasoning. arXiv preprint arXiv:2508.10057.Preprocessing NotesThis dataset contains minimally preprocessed raw EEG data:✓ Bad channels identified based on visual inspection✓ Trigger channel identified and annotated✗ No rereferencing (e.g. not yet rereferenced to average)✗ No filtering applied✗ No ICA artifact correction applied✗ No epoching appliedData Access and UseThis dataset is intended for research purposes.Code AvailabilityCode for data collection and analysis is available at https://github.com/chris-pinier/abstract_reasoningContactFor inquiries regarding this dataset, please contact:Michael D. NunezUniversity of Amsterdam m.d.nunez@uva.nlORChristopher PinierUniversity of Amsterdam c.pinier@uva.nl