Dual-Confidence Contrastive Decoding for Retrieval-Augmented Generation

2026-07-01

AI Summary

This work addresses the challenge of inaccurate responses in multi-document retrieval-augmented generation (RAG) caused by conflicting, noisy, or outdated information in retrieved contexts. The authors propose a training-free dual-confidence contrastive decoding method that jointly leverages document-level confidence—assessing the sufficiency of supporting documents—and token-level confidence—evaluating the reliability of next-token predictions. By dynamically weighting and contrasting positive and negative generation streams, the approach enables source-aware, reliable decoding. As the first study to explicitly tackle internal conflicts in multi-document RAG, this work also introduces DRQA, a new evaluation benchmark tailored to enterprise deep research scenarios. Experiments demonstrate that the proposed method achieves state-of-the-art average performance across both DRQA and standard multi-document question answering benchmarks, with particularly significant gains on DRQA, confirming the effectiveness of the proposed mechanism.

Abstract

Retrieval-augmented generation (RAG) increasingly requires models to answer questions from multiple retrieved documents, where only some sources are relevant and the retrieved bundle may contain stale, noisy, or conflicting evidence. Existing contrastive decoding methods primarily focus on resolving conflicts between the model's internal memory and the retrieved context. In contrast, we study the complementary problem of intra-context conflict in multi-document RAG. To evaluate this setting, we introduce DRQA, a factual-conflict question answering benchmark derived from enterprise deep-research scenarios, where answers are grounded in synthetic enterprise-specific facts that are designed not to be recoverable from the model's internal memory. We further propose Dual-Confidence Contrastive Decoding (DCCD), a training-free decoding method that combines document-level confidence, which estimates whether a document appears sufficient for answering the question, with token-level confidence, which estimates whether that document supports a confident next-token prediction. DCCD selects positive and negative document-conditioned streams using these dual-confidence signals and scales a document-level contrast by their confidence margin. Across DRQA and standard multi-document QA benchmarks, DCCD achieves the best average performance among full-context and contrastive decoding baselines, with the largest gains on DRQA. These results highlight the importance of source-aware, confidence-gated decoding when retrieved evidence is internally conflicting.

Problem

• retrieval-augmented generation
• multi-document RAG
• intra-context conflict
• factual-conflict question answering
• conflicting evidence

Innovation

• Dual-Confidence Contrastive Decoding
• Retrieval-Augmented Generation
• Intra-context Conflict
• Confidence-Gated Decoding
• Multi-document QA