چارچوب تشخيص اشتباه در پاسخ‌ مدل‌ زباني بزرگ در وظايف استدلالي

: Error Detection , Large Language Models , Logical Reasoning , Response eva‎luation , , Bias in eva‎luation , Calibration, , Jury Module , Weighted Voting

A Framework for Mistake Detection in Large Language Models Responses to Reasoning Tasks

مهندسي نرم افزار

Today, given the volume of data in large language models (LLMs) an‎d the likelihood of errors such as hallucination an‎d bias in the responses they generate, identifying mistakes in their outputs has become a key challenge. This challenge is especially critical in reasoning tasks that deman‎d high precision. Accordingly, we first examined the different types of reasoning in LLMs an‎d, focusing on one of the most widely used forms of reasoning, identified an‎d analyzed its common errors. We then eva‎luated methods for automatically detecting these errors. Although employing human judges is among the most credible ways to identify such errors, this approach is costly an‎d time-consuming an‎d suffers from limitations such as judge bias. One approach to addressing this problem is to use a large language model as a judge³. To make this approach more effective, several frameworks have been developed to provide accurate, highly scalable, an‎d relatively impartial assessments of responses with minimal implementation overhead. In this study, we present a multi-agent, LLM-based framework organized across multiple layers for the automatic detection of mistakes in LLM responses. The framework’s main innovations are examined in three key components. The first is a debate table that enables multidimensional analysis of a response by a set of intelligent agents using specialized pro‎mp‎ts. The second models the reasoning process with a dynamic graph to record an‎d critique claims. The third component performs intelligent aggregation of judgments: through a jury mechanism, it weights the judges’ opinions using a calibration model an‎d controls potential biases. eva‎luation on the HotpotQA dataset shows that, with this framework, accuracy increases from about 77.5% for baseline models to 89.87%. An ablation study further showed that removing agents leads to declines in accuracy an‎d soundness. Compared with the baselines, these results indicate that the simultaneous use of diverse agents, calibrated weighted aggregation, an‎d bias monitoring improves the reliability an‎d accuracy of mistake detection relative to comparable solutions. Moreover, for multi-step questions, the dynamic reasoning graph played an important role in surfacing structured logical reasoning patterns an‎d intra-text inconsistencies. Finally, the indicators computed during eva‎luation will reduce the system’s sensitivity to biases in future use.

6

151650