Toward explainable traffic crash liability analysis via neural-symbolic multimodal reasoning: Integrating visual perception with legal knowledge graphs

Determining traffic crash liability relies on rigorous on-site evidence and legal provisions and involves substantial manual intervention and potential human error. However, intelligent traffic monitoring systems focus on detection and tracking and cannot provide structured, rule-grounded decision support for formal crash-liability analysis. Moreover, emerging multimodal large language models (MLLMs) suffer from object hallucination that distorts or omits physical evidence, and their opaque reasoning cannot prod…

By the Truvace deskOctober 15, 20262 min read1 sourcePeer-reviewedTRV-2026-1404· cite & verify

Toward explainable traffic crash liability analysis via neural-symbolic multimodal reasoning: Integrating visual perception with legal knowledge graphs — "Motorway Traffic Jam" by dani_tic is licensed under CC BY-SA 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by-sa/2.0/.

In brief

Researchers propose CrashLex, a neural-symbolic framework for preliminary traffic crash liability assessment. It combines a vision-language model that extracts facts from crash scenes with a legal knowledge graph that maps those facts to statutory elements. A logic-driven protocol then applies the legal rules to the extracted facts.

The work targets a gap: current traffic monitoring detects and tracks but cannot support formal liability decisions, and general multimodal language models can hallucinate objects and give opaque reasoning. The reported results come from one held-out benchmark, and the system is framed as support for human review, so real-world performance and legal acceptance remain untested in the source.

Main points

The framework uses a dual-query vision-language design: Scene Queries model global context and Perception Queries extract fine-grained atomic features.
A domain-specific legal knowledge graph links observable physical conditions to statutory elements.
Legal Accuracy reached 94.0% under a fixed legal-output evaluation protocol on the held-out DeepAccident benchmark.
Object Hallucination Rate was 2.8%, versus 15.6% for the best-performing multimodal language model baseline, an 82.1% relative reduction.
Outputs rest on deterministic execution of predefined legal rules rather than only unconstrained probabilistic generation.

The gain

On the DeepAccident benchmark, grounding crash analysis in predefined legal rules cut object hallucination to 2.8% and reached 94.0% legal accuracy, giving reviewers traceable evidence-to-rule chains.

The rundown

The paper says crash liability determination depends on on-site evidence and legal provisions, involves substantial manual intervention, and carries the risk of human error. Existing intelligent traffic monitoring does not produce structured, rule-grounded decision support.

CrashLex addresses the hallucination problem with parallel semantic and symbolic streams, so that physical evidence is less likely to be distorted or omitted. Extracted facts are then tied to statutory provisions through the knowledge graph.

The aggregation protocol imposes explicit constraints from the extracted facts and statutory provisions on inference, enabling traceable and interpretable preliminary assessments. The authors present the output as verifiable reasoning chains for human review, not as a final legal decision.