Information Flow Tracking Quantifies Human Reasoning Dynamics, Mapping Behaviors Within a Single Metric Space

Human reasoning, a cornerstone of cognition, has long presented a challenge to researchers across multiple disciplines, including psychology, philosophy, and artificial intelligence. Qiguang Chen, Jinhao Liu, and Libo Qin, alongside their colleagues, now present a new method for understanding how people accumulate and transform information during reasoning. Their work introduces Information Flow Tracking, a technique that leverages large language models to quantify information entropy and gain at each step of the reasoning process. This approach successfully maps the universal landscape of human reasoning behaviours within a single, measurable space, capturing essential features, identifying systematic errors, and characterizing individual differences, and ultimately offering a quantitative link between theoretical models and observed reasoning processes.

Collective Behaviour and Problem Solving Research

This research consolidates the contributions of numerous scientists investigating collective behaviour and problem-solving strategies. The work draws upon expertise from diverse fields, including artificial intelligence, cognitive science, and animal behaviour, to advance understanding of complex cognitive processes. This collaborative effort aims to unravel the mechanisms underlying both individual and collective intelligence, paving the way for innovative solutions in various domains. The study builds upon datasets and methodologies developed by researchers across multiple institutions, fostering a comprehensive approach to understanding problem-solving dynamics.

This interdisciplinary collaboration leverages the strengths of each contributing scientist, resulting in a more nuanced and robust understanding of cognitive processes. This ongoing research continues to explore the interplay between individual cognition and collective intelligence, seeking to identify the key factors that contribute to successful problem-solving in both human and artificial systems. This pioneering technique tracks the development of uncertainty and cognitive effort at each step of a reasoning process, providing a detailed map of human thought. Researchers employed these models to estimate the probability of each reasoning step, generating quantifiable metrics for both uncertainty and the mental effort involved. Experiments involved presenting participants with reasoning tasks and then using the language model-based encoder to quantify the uncertainty and cognitive effort associated with each statement.

The resulting data revealed specific patterns; initial statements led to higher uncertainty, which decreased as the reasoning progressed. This allowed the team to establish a theoretical framework based on an information phase space, modelling reasoning as a trajectory within a two-dimensional system defined by uncertainty and cognitive effort. This formulation allows researchers to visualize and analyze how thought evolves from states of high uncertainty and low effort to those of low uncertainty and high effort. This work establishes a quantitative link between theoretical models of reasoning and measurable cognitive dynamics, revealing the underlying architecture of human thought processes. The research demonstrates that reasoning maintains a conserved structure, manifesting in smooth, continuous trajectories where thought evolves without loss of information. Experiments reveal that IF-Track accurately maps reasoning steps to a normalized phase space, creating an “information phase space” where arrows represent consistent flow direction.

This approach maintains coherent progression and interpretability, contrasting with disordered dynamics observed in non-reasoning scenarios. Measurements confirm that uncertainty decreases as intermediate conclusions accumulate, with a slight rebound at the final step, while cognitive effort rises steadily throughout the reasoning process. Analysis of local divergence along reasoning trajectories shows extended regions of near-zero divergence, consistent with approximately volume-preserving flow in phase space. Further investigations demonstrate IF-Track’s ability to distinguish between classical reasoning types, deductive, inductive, and abductive, via distinct trajectory patterns. By leveraging large language models as probabilistic encoders, the team successfully quantified information entropy and gain at each step of the reasoning process, establishing a metric space that captures universal patterns in how people think. This approach not only identifies essential features of reasoning, but also characterizes systematic errors and individual differences in cognitive style. The findings demonstrate that human reasoning can be modelled as an approximately incompressible information flow, aligning with principles observed in physical systems.

Specifically, the research reveals a dynamic interplay between uncertainty and cognitive effort, suggesting these variables act as conjugate pairs while approximately conserving phase-space volume. Furthermore, the method successfully distinguishes between classical reasoning types, deductive, inductive, and abductive, through unique trajectory patterns, and identifies reasoning errors as deviations from these typical patterns. Future work could explore the impact of individual cognitive biases and emotional states on reasoning trajectories, and investigate how this framework can be applied to improve artificial intelligence systems. This work establishes a quantitative link between theory and measurement, offering mechanistic insights into the architecture of reasoning and opening new avenues for understanding the human mind.