Researchers propose Dual-Confidence Contrastive Decoding (DCCD), a training-free method designed to address intra-context conflicts in retrieval-augmented generation where retrieved documents contain stale, noisy, or conflicting evidence. The approach combines document-level confidence, which estimates if a source is sufficient for answering, with token-level confidence, which evaluates support for next-token predictions.
- DCCD selects positive and negative document-conditioned streams using these dual-confidence signals and scales the contrast by their margin.
- To evaluate this setting, the authors introduce DRQA, a factual-conflict question answering benchmark derived from enterprise deep-research scenarios with synthetic facts.
- Across DRQA and standard multi-document QA benchmarks, DCCD achieves the best average performance among full-context and contrastive decoding baselines.
These results highlight the importance of source-aware, confidence-gated decoding when retrieved evidence is internally conflicting.