Overview
Paper: *Position: Logical Soundness is not a Reliable Criterion for Neurosymbolic Fact-Checking with LLMs*
Authors: Jason Chan, Robert Gaizauskas, Zhixue Zhao
Published: April 2025
arXiv: 2503.1385
Summary
As large language models (LLMs) are increasingly integrated into fact-checking pipelines, formal logic is often proposed as a rigorous means of mitigating bias, errors, and hallucinations in these models' outputs. For example, some neurosymbolic systems verify claims by using LLMs to translate natural language into logical formulae and then checking whether the proposed claims are logically sound — that is, whether they can be validly derived from premises verified to be true.
The authors argue that such approaches structurally fail to detect misleading claims due to systematic divergences between conclusions that are logically sound and inferences that humans typically make and accept.
Drawing on studies in cognitive science and pragmatics, the paper presents a series of cases in which logically sound conclusions systematically trigger human inferences that are not supported by the underlying premises. Because of this gap, a system that certifies claims purely on formal derivability can inadvertently greenlight statements that mislead human readers.
Proposed Alternative
The authors advocate a complementary approach: rather than treating LLMs' human-like reasoning tendencies as a flaw to be constrained by formal logic, these tendencies should be exploited as a feature. Specifically, LLMs can be used to audit the outputs of the formal components in neurosymbolic fact-checking systems, catching potentially misleading conclusions that logical soundness checks alone would miss.
Key Points
- Formal logical soundness and human acceptability of inferences systematically diverge.
- Neurosymbolic pipelines that rely solely on logical validity checks can fail to flag misleading claims.
- Evidence from cognitive science and pragmatics motivates a set of illustrative failure cases.
- LLMs' human-like reasoning biases can serve as a detection mechanism for logically valid but pragmatically misleading conclusions.