Abstract
Human reasoning is inherently non-monotonic, often involving the retraction of previously held conclusions upon the receipt of new, conflicting information. This study investigates the cognitive adequacy of five prominent non-monotonic reasoning (NMR) frameworks—System P, Weak Completion Semantics (WCS), Reiter’s Default Logic, System Z, and c-representations—by evaluating their performance against the "Suppression Task" benchmark from cognitive psychology. Our findings indicate that most standard NMR systems fail to replicate human-like inference suppression unless specifically tuned. Notably, Weak Completion Semantics, utilizing three-valued logic and abnormality predicates, demonstrates the highest alignment with empirical human data.
Introduction: The Challenge of Non-Monotonicity
In classical logic, monotonicity dictates that the addition of new information cannot invalidate existing truths. However, human "commonsense" reasoning frequently violates this principle. For instance, given the premise "If she has an essay, she studies late in the library," an individual might conclude she is studying late upon learning she has an essay. Yet, if subsequently informed that "The library is closed," the initial conclusion is retracted. This phenomenon, known as the Suppression Task, was famously documented by Ruth Byrne in 1989. While AI researchers have developed various Non-Monotonic Logics (NMLs) to handle such dynamics, their alignment with actual human cognitive processes remains a critical area of inquiry. This paper seeks to identify which formalisms are "cognitively adequate"—capable of drawing the same conclusions as human subjects.
The Suppression Task and Experimental Benchmarks
The study utilizes experimental results from Byrne (1989) as a primary benchmark. In the baseline case (α), approximately 98% of human participants performed the Modus Ponens inference. However, in the suppression case (γ), where a necessary condition is introduced (e.g., "If the library stays open..."), the rate of inference dropped significantly to 38%.
To further probe the formal models, the authors introduced variations (γ' and γ'') to test how different interpretations of the additional premise—such as whether the library is known to be open or closed—impact the logical output.
Methodology: Evaluating AI Reasoning Frameworks
Five "de-facto standard" NMR systems were evaluated:
- System P: Based on "typicality" and preferential reasoning.
- Weak Completion Semantics (WCS): A logic programming approach employing three-valued logic (True, False, Unknown).
- Reiter’s Default Logic: Utilizes default rules to fill information gaps.
- System Z: A ranking-based system measuring world "implausibility."
- c-representations: An advanced ranking system assigning individual impact weights to different conditional statements.
The researchers translated the natural language premises of the Suppression Task into formal knowledge bases for each system to determine if they would "suppress" the inference when the interfering premise was introduced.
Analysis and Results
The analysis revealed a significant discrepancy between standard AI logic and human behavior. Most prominent systems—System P, System Z, and Reiter’s Default Logic—failed to replicate the suppression effect in their standard configurations, proving "too logical" by maintaining the initial conclusion despite the interfering information.
In contrast, Weak Completion Semantics (WCS) was uniquely successful. By allowing for an "Unknown" state and employing "abnormality predicates" (e.g., "She studies late unless something is abnormal"), WCS naturally mirrored the human drop in confidence. Furthermore, the authors discovered that while System Z and c-representations could be adapted to work, they required modeling that mimicked the structure of WCS. This suggests that problem representation is as crucial as the underlying logical engine. A formal proof (Theorem 2) further established that WCS operates on fundamentally different principles than the other four systems, explaining its superior cognitive alignment.
Conclusion and Future Directions
The success of WCS highlights that human reasoning treats conditionals not as simple mathematical rules, but as "licenses for implication" laden with hidden assumptions. While this study provides a rigorous comparison for the Suppression Task, future work must address more complex, multi-step reasoning and the nuances of natural language interpretation. Exploring "conceptual adequacy"—how humans translate linguistic input into internal mental models—and the temporal order of information processing remains vital for developing AI that truly simulates human thought.
Summary updated