Reconfiguring Context for LLM-Based Auditing: A Theoretical Framework for Evidence Alignment and Traceability
Reconfiguring Context for LLM-Based Auditing: A Theoretical Framework for Evidence Alignment and Traceability
Seyyed mohammad Abbasnia1
1) Assistant Professor of Manegement& Acounting, Payame Noor University, Tehran, Iran
IOI Code
:
AAHD-ABKCCF
Publication :
National Conference on Information Technology, Nanotechnology, Artificial Intelligence and Technological Futures Studies(itnaf.ir)
Abstract :
Large Language Models (LLMs) are increasingly deployed in environments characterized by lengthy texts and dispersed information. In the auditing of financial statements, however, the central challenge is not merely text retrieval or summarization, but the reconfiguration of context such that relevant evidence is meaningfully juxtaposed and the “claim–evidence–conclusion” chain of reasoning is constructed in a traceable and defensible manner. Adopting a purely theoretical perspective, this paper reframes context reconfiguration as a design problem in LLM-based systems for auditing. The core argument is that two structural features of the reporting environment—disclosure dispersion and informational noise—undermine evidence alignment by increasing extraneous cognitive load and disrupting attentional allocation, thereby jeopardizing the defensibility of professional judgment under institutional requirements for quality and accountability. Building on this premise, the paper proposes a conceptual model and a set of theoretical propositions that explain the relationships among disclosure structure, cognitive load, and the quality of evidence organization. These propositions are then translated into design principles for LLM-based systems, including goal-oriented noise gating, faithful and referenceable compression, and claim–evidence–conclusion mapping such that each assertion is explicitly linked to its corresponding textual source and remains traceable for independent review. The paper further draws a fundamental distinction between using LLMs as tools for “text generation” and employing them as mechanisms for “evidence-context organization,” demonstrating that only the latter is consistent with the logic of defensibility and quality in auditing. Finally, a research agenda for the artificial intelligence and machine learning literature is proposed, emphasizing construct definition for context-centric concepts and the development of evaluation metrics for context reconfiguration in assurance applications.
Keywords : large language models context reconfiguration informational noise disclosure dispersion cognitive load llm based auditing