VeriGraph: Towards Verifiable Data-Analytic Agents

VeriGraph is a newly proposed neuro-symbolic reasoning framework designed to enhance the verifiability of LLM-based agents in data-analytic tasks by constructing explicit evidence-directed acyclic graphs (DAGs) during execution. It introduces three evidence-expansion primitives—computational, grounding, and derivational expansion—to unify raw data, interpreter variables, and natural-language claims, resulting in improved traceability and semantic support for claims. The model, VeriGraph-8B, demonstrated superior performance on four benchmarks, achieving an 87.61% Grounding Rate, indicating its potential for creating auditable evidence graphs in AI applications.