Neuro-Symbolic Rule Discovery: Empowering LLMs with Causality for Vehicle Diagnostics

Hugo Math, Julian Lorenz, Rainer Lienhart

Track: tiny / short paper (up to 4 pages)

Keywords: causal discovery, event sequences, llms, autoregressive models, stochastic systems, automotive

TL;DR: CAREP is a neuro-symbolic framework that automates vehicle fault diagnosis by constraining Large Language Models with causal discovery, enabling the explainable synthesis of rigorous Boolean error patterns from noisy, high-dimensional event streams.

Abstract: Defining Boolean logic for vehicle fault detection of error patterns (EPs) is a manual, error-prone bottleneck process in automotive safety. Standard LLMs struggle to automate this task, as they prioritize semantic plausibility over logical necessity. We propose CAREP, a framework that empowers LLMs with causal discovery to extract strict diagnostic rules from noisy high-dimensional event sequences. Instead of relying on semantic correlations, CAREP provides the LLM with a grounded set of causal drivers (excitatory) and constraints (inhibitory). This enables the automated synthesis of accurate, human-readable rules alongside reasoning traces. On a real-world dataset of 29,100 unique codes, CAREP achieves superior rule reconstruction accuracy compared to standard RAG baselines.

Anonymization: This submission has been anonymized for double-blind review via the removal of identifying information such as names, affiliations, and identifying URLs.

Funding: Yes, the presenting author of this submission falls under ICLR’s funding aims, and funding would significantly impact their ability to attend the workshop in person.

Submission Number: 21