ARGUS: Feedback-Reinforced Gradual LLM-Based Framework for Interpretable and Robust Archive Review

ACL ARR 2025 May Submission3687 Authors

Abstract:

Automated archive review faces challenges in interpreting domain-specific semantics and ensuring traceable decisions, because existing methods relying on rigid rules or generic language models lack complex context understanding and review transparency.
Regarding these issues, we propose ARGUS, a feedback-reinforced gradual framework for archive review.
ARGUS uses hierarchical rule-embedded prompts for stepwise inference, feedback-driven sample enhancement via LLM inference logs for robustness, and parameter-efficient fine-tuning via low-rank adaptation.
Evaluations on real-world archives and benchmarks show ARGUS achieves 10.5–15.5\% higher accuracy than baselines, reduces ASR by 25\%, and has been proven to effectively complete review tasks under limited resources.

Paper Type: Long

Research Area: NLP Applications

Research Area Keywords: Archive Review，LLM-based，Hierarchical Rule-Embedded Prompting，Feedback-Reinforced Learning

Languages Studied: Chinese，English

Submission Number: 3687