Probabilistic Inconsistency Framework for Neuro-Symbolic QA Challenges LLM Temporal Reasoning Deficit
A recent study published on arXiv disputes the common belief that large language models (LLMs) struggle with temporal reasoning due to limitations in autoregressive logical deduction. The authors propose that the real issue lies in the unstructured representation of text-to-event. They present a neuro-symbolic framework for question-answering that utilizes a Probabilistic Inconsistency Signal (PIS) to differentiate perceptual mistakes from reasoning errors. This architecture transforms unstructured text into clear event graphs and interval constraints, separating semantic extraction from a symbolic reasoning system. The PIS integrates symbolic credal intervals with epistemic neural uncertainty from Evidential Deep Learning to identify structural breaks. The paper can be found at arXiv:2605.04243.
Key facts
- Paper challenges the narrative that temporal reasoning is the fundamental bottleneck in LLMs.
- Locus of failure identified as unstructured text-to-event representation.
- Introduces Probabilistic Inconsistency Signal (PIS) to isolate perceptual errors from reasoning failures.
- Architecture decouples semantic extraction from symbolic reasoning.
- Uses explicit event graphs and interval constraints.
- PIS unifies symbolic credal intervals with epistemic neural uncertainty.
- Published on arXiv with ID 2605.04243.
- Announcement type is new.
Entities
Institutions
- arXiv