StitchRAG compresses retrieved passages into compact contexts using diffusion stitching
An arXiv paper proposes a retrieval-augmented generation framework that trims retrieved passages to a per-passage size and stitches selected sentences into a coherent context.
In plain words
Retrieval-augmented generation (RAG) is a technique where a language model answers questions using documents fetched from an external source. The StitchRAG paper describes a way to shrink those fetched documents so the model keeps the useful evidence without processing long, noisy text. It selects key sentences and uses a text diffusion method to join them into one smooth context.
What was announced
The authors frame the problem as a trade-off: long retrieved contexts add computational overhead and can degrade performance when they contain noisy or irrelevant content. StitchRAG addresses this by formulating a utility maximization that balances retained information against computation costs, setting a size reduction ratio for each passage individually. The system then extracts answer candidates and query-related keywords from passages as anchors, and collects sentences containing those anchors under the derived size constraint.
To make the compressed output readable, the framework connects the collected sentences using text diffusion, producing what the authors describe as a coherent and natural context. On five open-domain question answering benchmarks, the paper reports that StitchRAG consistently outperformed diverse multi-step and compression baselines. Ablation studies attributed the gains to the dynamic size allocation, anchor preservation, and diffusion-based stitching components. The evidence is limited to the paper's abstract, so benchmark details, model configurations, and comparison numbers are not available in the supplied material.
What it does not show
- Only the paper's abstract was supplied; benchmark names, quantitative results, and experimental setup are not available to verify the reported outperformance.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.