[Paper Review] Rowen: Adaptive Retrieval-Augmented Generation for Hallucination Mitigation in LLMs

Hanxing Ding, Liang Pang+3 more|arXiv (Cornell University)|Feb 16, 2024

Epilepsy research and treatment4 citations

TL;DR

Rowen is a novel, adaptive retrieval-augmented generation framework that mitigates hallucinations in large language models (LLMs) by dynamically activating external retrieval only when internal reasoning is inconsistent across multilingual perturbations. It achieves state-of-the-art hallucination mitigation with significantly fewer retrieval calls than existing methods.

ABSTRACT

Hallucinations present a significant challenge for large language models (LLMs). The utilization of parametric knowledge in generating factual content is constrained by the limited knowledge of LLMs, potentially resulting in internal hallucinations. While incorporating external information can help fill knowledge gaps, it also introduces the risk of irrelevant information, thereby increasing the likelihood of external hallucinations. To balance the use of parametric knowledge within LLMs and external information, in this study, we present Rowen, a novel framework that enhances LLMs with an adaptive retrieval augmentation process tailored to address hallucinated outputs. Rowen introduces a consistency-based hallucination detection module, which assesses the model's uncertainty regarding the input query by evaluating the semantic inconsistencies in various responses generated across different languages or models. When high uncertainties in the responses are detected, Rowen activates the retrieval of external information to rectify the model outputs. Through comprehensive empirical experiments, we demonstrate that Rowen surpasses the current state-of-the-art in both detecting and mitigating hallucinated content within the outputs of LLMs.

Motivation & Objective

• To address the dual challenge of internal hallucinations (due to limited parametric knowledge) and external hallucinations (from irrelevant retrieval) in LLMs.
• To develop a method that selectively activates retrieval only when necessary, minimizing unnecessary external evidence injection.
• To improve factual consistency and reliability of LLM outputs through a multilingual semantic-aware detection mechanism.
• To balance factual accuracy with inference efficiency by reducing redundant retrieval calls.
• To enable a harmonious integration of internal parametric knowledge and external evidence in LLM reasoning.

Proposed method

• Rowen employs a multilingual semantic-aware hallucination detection module that perturbs a given query into multiple semantically equivalent versions across languages.
• It evaluates response consistency across these perturbed queries in different languages; inconsistency signals potential internal hallucination.
• When inconsistency is detected, Rowen triggers a retrieval-augmented generation process to correct the output using relevant external evidence.
• If responses remain consistent across languages, the original internal reasoning is preserved, avoiding unnecessary retrieval.
• The system uses a threshold-based consistency metric to determine when retrieval is needed, balancing detection sensitivity and retrieval cost.
• The method is designed to minimize both internal and external hallucinations by activating retrieval only when internal reasoning is unreliable.

Experimental results

Research questions

• RQ1 RQ1: Can a multilingual consistency check effectively detect internal hallucinations in LLMs?
• RQ2 RQ2: How does adaptive retrieval based on consistency detection improve factual accuracy compared to always-on retrieval?
• RQ3 RQ3: What is the optimal threshold for consistency-based retrieval activation to balance detection and retrieval overhead?
• RQ4 RQ4: How does the number of perturbed questions affect detection performance and computational cost?
• RQ5 RQ5: How does Rowen compare to existing RAG methods in terms of retrieval efficiency and hallucination mitigation?

Key findings

• Rowen achieves state-of-the-art hallucination mitigation performance on both TruthfulQA and StrategyQA benchmarks.
• Rowen reduces the average number of retrieval calls per question to 1.5 on TruthfulQA and 0.5 on StrategyQA, outperforming FLARE (2.1 and 3.9) and Detect-and-Mitigate (7.2 and 5.5).
• A consistency threshold of 0.6 yields optimal performance, balancing detection accuracy and risk of external hallucination from irrelevant retrieval.
• Using six perturbed questions achieves near-optimal performance with diminishing returns beyond that, enabling efficient deployment.
• The method significantly reduces the risk of external hallucination by avoiding retrieval when internal reasoning is consistent and reliable.
• Rowen’s adaptive mechanism results in fewer retrieval calls than all compared methods while maintaining or improving factual accuracy.