PyRAG: Multi-Hop RAG as Executable Code
A new framework called PyRAG reformulates multi-hop retrieval-augmented generation (RAG) as program synthesis and execution. The approach addresses brittleness in existing RAG systems on multi-hop questions by representing reasoning steps as executable code rather than free-form natural language. This method leverages code-specialized language models to produce structured, verifiable reasoning trajectories, reducing query drift and improving error detection. The paper is published on arXiv under ID 2605.12975.
Key facts
- PyRAG reformulates multi-hop RAG as program synthesis and execution.
- Existing RAG systems are brittle on multi-hop questions.
- Current methods use free-form natural language for reasoning.
- Free-form reasoning leads to implicit intermediate states and query drift.
- Self-reflection is unreliable when done by the same model.
- Multi-hop QA is a form of step-by-step computation.
- Code-specialized LMs are trained to operate on structured processes.
- The paper is on arXiv with ID 2605.12975.
Entities
Institutions
- arXiv