Interactive DualChecker for Mitigating Hallucinations in Distilling Large Language Models

TL;DR

DualChecker addresses hallucinations in LLM-based knowledge distillation by coupling ContextAligner with an interactive confidence-based checking loop to align outputs with human labeling standards. It retrieves similar contexts, generates rationales, and iteratively refines teaching templates through teacher and student feedback, improving both teacher and student performance in a challenging green-innovation domain. Empirical results show significant F1 gains across classification, causality extraction, and pathway identification tasks, with up to 17% improvement for teachers and 10% for students, and the approach remains effective with both white-box and black-box LLMs. The method is open-sourced, enabling reproducibility and further research in domain-specific distillation where annotated data are scarce and domain standards are strict.

Abstract

Large Language Models (LLMs) have demonstrated exceptional capabilities across various machine learning (ML) tasks. Given the high costs of creating annotated datasets for supervised learning, LLMs offer a valuable alternative by enabling effective few-shot in-context learning. However, these models can produce hallucinations, particularly in domains with incomplete knowledge. Additionally, current methods for knowledge distillation using LLMs often struggle to enhance the effectiveness of both teacher and student models. To address these challenges, we introduce DualChecker, an innovative framework designed to mitigate hallucinations and improve the performance of both teacher and student models during knowledge distillation. DualChecker employs ContextAligner to ensure that the context provided by teacher models aligns with human labeling standards. It also features a dynamic checker system that enhances model interaction: one component re-prompts teacher models with more detailed content when they show low confidence, and another identifies borderline cases from student models to refine the teaching templates. This interactive process promotes continuous improvement and effective knowledge transfer between the models. We evaluate DualChecker using a green innovation textual dataset that includes binary, multiclass, and token classification tasks. The experimental results show that DualChecker significantly outperforms existing state-of-the-art methods, achieving up to a 17% improvement in F1 score for teacher models and 10% for student models. Notably, student models fine-tuned with LLM predictions perform comparably to those fine-tuned with actual data, even in a challenging domain. We make all datasets, models, and code from this research publicly available.

Figures (3)

• Figure 1: Distribution of Hallucination Types in a Preliminary Experiment.
• Figure 2: DualChecker: An Interactive Approach to Mitigate Hallucinations and Enhance Knowledge Distillation of LLMs. The process includes (1) ContextAligner, which retrieves similar data and guides prompting using LLM-generated rationales; (2) the LLM then generates a confidence score, rationale, and prediction, which a Checker evaluates against a threshold to determine the need for re-prompting; (3) if the confidence score meets the threshold, the prediction fine-tunes the student model, which outputs predictions and probabilities; (4) a second Checker examines these probabilities, identifies the least confident case, generates a rationale, and updates the teaching template for subsequent prompting.
• Figure 3: Results with Varying Numbers of Shots using GPT-3.5 Turbo.