Abstract
Traditional belief revision frameworks often rely on the principle of minimalism, which advocates minimal changes to existing beliefs. However, research in human cognition suggests that people are inherently driven to seek explanations for inconsistencies, thereby striving for explanatory understanding rather than minimal changes when revising beliefs. Traditional frameworks often fail to account for these cognitive patterns, relying instead on formal principles that may not reflect actual human reasoning. To address this gap, we introduce Human-Aware Belief Revision, a cognitively-inspired framework for modeling human belief revision dynamics, where given a human model and an explanation for an explanandum, revises the model in a non-minimal way that aligns with human cognition. Finally, we conduct two human-subject studies to empirically evaluate our framework under real-world scenarios. Our findings support our hypotheses and provide insights into the strategies people employ when resolving inconsistencies, offering some guidance for developing more effective human-aware AI systems.
Introduction
In the field of Artificial Intelligence, belief revision is the process by which an agent updates its knowledge base upon encountering information that contradicts its current beliefs. For decades, the Alchourrón, Gärdenfors, and Makinson (AGM) framework has served as the standard, predicated on the principle of minimalism: an agent should modify its existing beliefs as little as possible to accommodate new data.
However, human cognition operates differently. When faced with a contradiction—such as learning that a penguin is a flightless bird despite believing all birds fly—humans do not merely excise the conflicting fact. Instead, they seek an explanation that reconciles the inconsistency, often leading to broader, non-minimal revisions of their mental models. This paper argues that AI systems must move beyond minimalist logic to effectively model and interact with human reasoning.
The Human-Aware Belief Revision Framework
To bridge formal logic and cognitive science, we propose the Human-Aware Belief Revision framework. This framework represents a human's belief base as a set of ground facts and general rules.
Central to this framework is the Explanation-Guided Revision Operator ($⊙$). Unlike traditional operators that seek the smallest set of beliefs to remove, the $⊙$ operator is guided by an explanation for the new, contradictory fact (the explanandum). It selects a "correction set" that restores consistency while ensuring the final belief base entails the provided explanation. This allows for "non-minimal" changes, such as discarding a general rule in favor of a more complex explanation that accounts for disabling conditions.
Empirical Evaluation
We conducted two human-subject studies via the Prolific platform to test the hypothesis that humans prioritize explanatory understanding over minimalism.
Experiment 1: Explanation Generation
Participants (n=62) were presented with nine inconsistent scenarios (e.g., a match being struck but failing to produce light) and asked to provide explanations. These responses were coded to determine if they implied minimal or non-minimal belief revisions.
Experiment 2: Explanation-Guided Revision
Participants (n=60) were provided with the same scenarios along with a plausible explanation (e.g., "the match was wet"). They were then asked to revise their original beliefs based on this new information.
Results
The results strongly support the non-minimalist nature of human belief revision. In Experiment 1, 83.37% of participant-generated explanations implied non-minimal revisions. Rather than rejecting a single fact, participants sought complex explanations that challenged general rules.
In Experiment 2, when provided with an explanation, the preference for non-minimal revision increased to 92.33%. On average, participants changed 1.88 beliefs per scenario, significantly higher than the single change predicted by minimalist frameworks.
Statistical analysis using Wilcoxon tests across all categories yielded p-values near zero ($p \approx 1.01 \times 10^{-50}$), confirming that the drive for explanatory understanding is a robust feature of human cognition across various problem types.
Conclusion
Our findings demonstrate that humans are explanation-seekers rather than minimalists. For AI systems to be truly "human-aware," they must prioritize explanatory understanding in their belief revision processes. Future work will focus on integrating this framework into active AI systems, such as robots performing model reconciliation, and exploring how epistemic attitudes like skepticism influence revision dynamics.
Summary updated