SCIENTIFIC DEVELOPMENTS FOR IMPROVING MODERN EDUCATION

The increasing complexity and scale of modern corporate accounting have created an urgent demand for reliable, transparent, and regulation-compliant auditing systems.Although large language models (LLMs) demonstrate powerful reasoning abilities, their use in auditing tasks is hindered by challenges such as hallucinations, lack of transparent multi-step reasoning, and weak alignment with strict financial reporting standards.This paper introduces AuditGuard, a robust LLM-based framework intended to automate financial audit inquiries while providing explainable, standardsaware reasoning.AuditGuard merges chain-of-thought prompting with a Reasoning Trace Graph that systematically maps intermediate inferences to relevant accounting clauses, enabling explicit traceability and regulatory interpretability.To address hallucination risks, a rejection mechanism refined through reinforcement learning on expert annotations allows the model to abstain from issuing conclusions in cases of incomplete or ambiguous reasoning.Experiments on a synthetically generated yet realistic auditing dataset demonstrate that AuditGuard outperforms strong baselines in terms of compliance accuracy, explanation consistency, and hallucination rejection, suggesting its practical viability as a trustworthy assistant in financial audit workflows.