Neuro-symbolic constraint verification for LLM-driven internet finance transaction execution

Large language models (LLMs) struggle to ensure compliance with strict business rules governing financial transactions. To address this, the Verifier-Actor Neuro-Symbolic Framework (VA-NSF) has been developed. This dual-layer architecture separates natural language understanding from formal constraint enforcement for executing internet finance transactions.

The neural Actor component translates financial instructions into a typed, schema-grounded abstract syntax tree (AST) using grammar-constrained decoding. Meanwhile, the symbolic Verifier assesses each AST against Answer Set Programming (ASP) integrity constraints, providing a formal approval certificate or a structured violation report that triggers targeted repairs.

In extensive testing on FinTxBench, a curated dataset of 8,600 annotated financial instructions across three complexity levels, VA-NSF achieved an impressive 91.8% overall transaction accuracy. The macro-averaged accuracy was 94.7% for Simple tier, 91.2% for Compound tier, and 89.6% for Constrained tier. The system only exhibited a constraint violation rate of at most 0.3%.

VA-NSF outperformed four LLM baselines by 13 to 51 percentage points and reduced violations by a factor of 30. The system operates with a median latency of 399 ms, of which ASP verification accounts for only 24 ms.

Independent ablations confirm the contributions of each component, and cross-domain experiments in payment processing, procurement, payroll, and inventory modules showed at most a 5.8 percentage point degradation. The formal guarantees are contingent upon the fidelity of the encoded rule set. The evaluation utilized a curated benchmark with up to 80 rules.

However, there are limitations related to noisy real-world language, the scale of rule sets, and human-escalation overhead. This work is licensed under a Creative Commons Attribution 4.0 International License, allowing for use, sharing, adaptation, distribution, and reproduction under specified conditions.