Meta's RAGCompass routes RAG retrievers by reading passages, +28% answer F1 over best fixed retriever
_reachsumit · x · 2026-11-06
A Meta-authored paper, RAGCompass, tackles retriever routing in RAG: no retriever wins on all queries, but existing routers rely on query features and a fixed retriever pool, failing on unseen retrievers or corpora.
How RAGCompass works:
- It reads the retrieved passage sets themselves, so outputs from unfamiliar retrievers or corpora stay directly comparable — routing generalizes by design.
- A reranker is contrastively fine-tuned against the answer quality each evidence set produces, driving most of the gain; a small comparison head adds a moderate boost. Neither component has retriever-specific parameters.
- It improves answer F1 by 28% relative to the best fixed retriever, beating fusion baselines, three query-only routers, and a prompted 8B model.
- Generalization experiments: it keeps most of its gain on unseen corpus, retriever, or reader, while query-only baselines fall to or below the best fixed retriever.
More from Research
- Zhipu chief scientist reveals GLM pretraining data roadmap: GLM-5 at 28.5T tokens, 100T next — rickasaurus · 2026-11-06
- Laude Institute grants $10M to UChicago team building AI weather forecasts for farmers — ChenhaoTan · 2026-11-06
- Quick test: Gemini 3.1 Pro knows the answer in its weights, Flash doesn't — giffmana · 2026-11-06
- JupyterHealth team wins $10M Laude Institute Moonshot Prize for open-source health data infra — irenetrampoline · 2026-11-06
- SmoothScore: SOTA LLMs still 'smooth' nuanced research findings into overclaims — gregd_nlp · 2026-11-06
- Schmidhuber: Yale's fast weight programming course reprises his 1990s linear Transformer work — SchmidhuberAI · 2026-11-06