OpenDecoder: Open Large Language Model Decoding to Incorporate Document Quality in RAG

Retrieval-augmented generation typically assumes that retrieved passages are sufficiently relevant and useful, even though real retrieval results vary considerably in quality. OpenDecoder exposes the large language model decoding process to explicit document-quality indicators, including relevance scores, ranking scores, and query performance prediction signals. This design allows generation to respond directly to the reliability of the retrieved context and can be integrated flexibly with different post-training objectives and external indicators. Experiments on five benchmarks demonstrate stronger effectiveness and robustness than existing approaches under varying levels of retrieval noise.

📄 In Proceedings of the ACM Web Conference 2026 (TheWebConf 2026), pp. 2252–2262.
🔗 Paper Link (ACM Digital Library)