Intent-Preserving Neuro-Symbolic Repair for Safe LLM-Driven Financial Transactions (Journal of Computing and Electronic Information Management)

Summary

• LLM repair in a re-implemented VA-NSF loop produced intent-mismatched commits for 32.1% of requests across 296 evaluation instances.p.1
• On correctly parsed requests, 34.6% of VA-NSF's repaired commits changed a user-stated field.p.1
• Intent-Preserving Repair (IPR) raised safe task success from 60.5% to 82.4% (p < 10⁻¹⁹).p.1
• IPR reduced required LLM calls from 2.49 to 1.15 per request.p.1
• The SGD-Fin benchmark consists of 607 money-transfer episodes extracted from the Schema-Guided Dialogue corpus across nine business rules.p.2
• Symbolic stages of grounding, verification, and repair execute with a median time of 7.0 ms per request.p.6

Key insights

1. LLM repair in VA-NSF led to intent-mismatched commits in 32.1% of total requests.p.1
2. IPR raised safe task success from 60.5% to 82.4% and produced zero drifted repairs.p.1
3. An independent policy oracle agreed with the ASP verifier on all 5,463 AST and state pairs.p.3
4. The initial parse of the neural Actor was intent-faithful for 73.6% of the 296 evaluation instances.p.4
5. Removing the intent lock caused safe success to drop to 21.6% and unsafe commits to rise to 70.6%.p.5
6. Under currency rule C5, VA-NSF committed a payment in a non-requested currency in 96.7% of instances.p.5
7. The symbolic execution stages take a median time of 7.0 ms per request.p.6