Improving Contextual Faithfulness of Large Language Models via Retrieval Heads-Induced Optimization

2025-01-23

AI Summary

This paper addresses the issue of answer unfaithfulness to retrieved contexts in Retrieval-Augmented Generation (RAG) for Long-Form Question Answering (LFQA). Methodologically, it first identifies a strong correlation between retrieval attention heads and generation faithfulness; introduces a masked retrieval head mechanism to synthesize unfaithful samples; and proposes control tokens to steer contrastive decoding between faithful and unfaithful answers. It also establishes GroundBench—the first LFQA benchmark explicitly designed to evaluate context faithfulness. Experiments demonstrate that RHIO significantly improves faithfulness (+12.7% on GroundBench), outperforming GPT-4o while preserving answer quality and factual consistency. Key contributions include: (1) a retrieval-head-induced optimization mechanism, (2) a control-token-driven contrastive decoding paradigm, and (3) the GroundBench evaluation benchmark.

Technology Category

• Natural Language Processing: Question Answering
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: Trust and reliance of crowd workers and data experts on GenAI

Abstract

Ensuring contextual faithfulness in retrieval-augmented large language models (LLMs) is crucial for building trustworthy information-seeking systems, particularly in long-form question-answering (LFQA) scenarios. In this work, we identify a salient correlation between LFQA faithfulness and retrieval heads, a set of attention heads responsible for retrieving contextual information. Leveraging this insight, we propose RHIO, a framework designed to teach LLMs to explicitly discriminate between faithful and unfaithful generations. RHIO first augments unfaithful samples that simulate realistic model-intrinsic errors by selectively masking retrieval heads. Then, these samples are incorporated into joint training, enabling the model to distinguish unfaithful outputs from faithful ones conditioned on control tokens. Furthermore, these control tokens are leveraged to self-induce contrastive outputs, amplifying their difference through contrastive decoding. Additionally, to facilitate the evaluation of contextual faithfulness, we also introduce GroundBench, a comprehensive benchmark compiled from five existing LFQA datasets. Extensive experimental results on GroundBench demonstrate that RHIO significantly improves faithfulness, even outperforming GPT-4o.

Problem

• Large Language Models
• Improvement
• Faithful Response Generation

Innovation

• RHIO
• Contrastive Training
• Faithfulness Enhancement