Closing the Loop to Discover Psychological Theories with an Automated Cognitive Scientist

A team of researchers has introduced AutoCog, a fully autonomous agentic-AI system designed to close the loop of theory-building in cognitive science by independently proposing theories, designing experiments, and testing them against human behavior [1][2]. The system, described in a paper submitted to the arXiv preprint server on 24 June 2026, uses large-language-model agents to advocate competing theories expressed as executable cognitive models [1][2]. These agents design experiments to discriminate between the theories, collect behavioral data from participants recruited online, and score the theories based on their generative performance [2]. The cycle repeats as the system diagnoses failures and synthesizes improved successors, searching the space of theories, models, and experiments without manual intervention [2]. In the domain of decision-making, AutoCog recovered known strategies from simulated behavior, including unconventional ones, indicating its discoveries are driven by data rather than strictly bound by the priors of the underlying language models [2]. When tested with human participants, the system produced theories that outperformed the established theories it was seeded with and generalized to held-out studies across two different experimental settings [1][2]. It also surfaced a novel theory of multi-cue decision-making in which choices show diminishing sensitivity to feature values; the distinctive predictions of this theory were confirmed in a preregistered study with new participants [2]. The work addresses a longstanding bottleneck in cognitive science, where theory generation has remained a manual process even as data collection, modeling, and experiment design have been automated [2]. Problem solving in science often requires overcoming mental obstacles such as confirmation bias and functional fixedness, which are widely studied in psychology and cognitive sciences [5]. By automating the creative step of turning accumulated model failures into better theories, AutoCog demonstrates how an automated discovery system can make cognitive theory-building explicit, executable, and cumulative [2]. The paper was posted on arXiv, an open-access repository of electronic preprints that, as of November 2024, receives about 24,000 articles per month and is not peer reviewed [10]. The submission history lists Akshay Kumar Jagadish as the corresponding author, with the manuscript file sized at 1,490 KB [1]. The research sits within the broader context of embodied cognition, a group of theories investigating how cognition is shaped by bodily states and capacities, which challenges purely computationalist views of the mind [4]. While AutoCog operates in the domain of decision-making, its closed-loop architecture draws on principles from neural network design, where systems composed of interconnected layers learn hierarchical representations through iterative training [6].