S2G-RAG: Structured Sufficiency and Gap Judging for Iterative Retrieval-Augmented QA
A novel framework named S2G-RAG (Structured Sufficiency and Gap-judging RAG) has been introduced to enhance multi-hop question answering within retrieval-augmented generation systems. This framework features a dedicated controller, S2G-Judge, which assesses at each step whether the existing evidence memory can adequately answer the question. If it falls short, the system generates structured gap items that outline the lacking information, which are then incorporated into the subsequent retrieval query, ensuring consistent multi-turn retrieval paths. To minimize noise buildup, S2G-RAG utilizes a sentence-level Evidence Context by selecting a concise set of pertinent sentences. This method tackles prevalent challenges in iterative processes, such as responding with incomplete evidence or accumulating irrelevant text. The paper can be found on arXiv with the identifier 2604.23783.
Key facts
- S2G-RAG is an iterative framework for retrieval-augmented generation.
- It uses an explicit controller called S2G-Judge.
- S2G-Judge predicts sufficiency of evidence and outputs structured gap items.
- Gap items are mapped into the next retrieval query.
- The framework maintains a sentence-level Evidence Context.
- It addresses incomplete evidence chains and noise accumulation.
- The paper is on arXiv with ID 2604.23783.
Entities
Institutions
- arXiv