Research map / Large language models / Retrieval and knowledge
Retriever Portfolios: A Principled Approach to Adaptive RAG
Miltiadis Stouras, Vincent Cohen-Addad, Silvio Lattanzi, Ola Svensson
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| select the retriever portfolio greedily by best-of-k coverage instead of top-k by average training score (ablation) | retriever portfolio (k=5) · pooled 4 QA benchmarks | support recall: 0.492 -> 0.594 | Improved |
| select the retriever portfolio greedily by best-of-k coverage instead of top-k by average training score (ablation) | retriever portfolio (k=5) · pooled 4 QA benchmarks | support F1: 0.432 -> 0.500 | Improved |
| use a diverse portfolio instead of retrieving k times more documents with the single best retriever (ablation) | single best retriever · pooled 4 QA benchmarks | support F1: DS/Vendi 0.32 -> 0.11 as k 1 -> 5; portfolio 0.431 -> 0.500 | Improved |
| use routed all-pool portfolio (k=4, l=3) instead of best single-family retriever | Gemma-3-27B-It RAG · average EM over 4 QA datasets | exact match: 0.424 -> 0.464 | Improved |
| use routed all-pool portfolio (k=4, l=3) instead of nearest-neighbor retrieval | Llama-3.1-70B-Instruct RAG · average EM over 4 QA datasets | exact match: 0.403 -> 0.474 | Improved |
| increase routed portfolio members executed per query from l=2 to l=3 (ablation) | all-pool portfolio (k=4) · Gemma-3-27B-It, avg EM | exact match: 0.456 -> 0.464 | Improved |
| increase routed portfolio members executed per query from l=2 to l=3 (ablation) | all-pool portfolio (k=4) · TriviaQA, Llama-3.1-70B-Instruct | exact match: 0.7073 -> 0.6859 | Hurt |
| restrict portfolio to Vendi-only retrievers instead of full heterogeneous pool (ablation) | routed portfolio · HotpotQA, Gemma-3-27B-It | exact match: 0.5523 -> 0.3927 | Hurt |
| use fixed offline Vendi portfolio instead of Vendi-RAG inference-time tuning | Vendi-RAG · Gemma-3-27B-It, MPNet/Vendi | EM vs tokens and latency | Improved |
| use fixed offline Vendi portfolio instead of Vendi-RAG inference-time tuning | Vendi-RAG · Llama-3.1-70B-Instruct, MPNet/Vendi | EM vs tokens and latency | Mixed |
| add Vendi-RAG adaptive tuning to Gemma-3-27B-It RAG | Gemma-3-27B-It RAG · HotpotQA | exact match: 0.3945 -> 0.2853 (vs nearest-neighbor, T=20) | Hurt |
Closest papers by meaning
- Cooperative Retrieval-Augmented Generation for Question Answering: Mutual Information Exchange and Ranking by Contrasting Layers (NeurIPS 2025 · similarity 0.89)
- DynamicRAG: Leveraging Outputs of Large Language Model as Feedback for Dynamic Reranking in Retrieval-Augmented Generation (NeurIPS 2025 · similarity 0.88)
- Adapting to Non-Stationary Environments: Multi-Armed Bandit Enhanced Retrieval-Augmented Generation on Knowledge Graphs (AAAI 2025 · similarity 0.88)
- Speculative RAG: Enhancing Retrieval Augmented Generation through Drafting (ICLR 2025 · similarity 0.87)
- One Token Can Help! Learning Scalable and Pluggable Virtual Tokens for Retrieval-Augmented Large Language Models (AAAI 2025 · similarity 0.87)
- RAGGED: Towards Informed Design of Scalable and Stable RAG Systems (ICML 2025 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
