Research map / Large language models / Retrieval and knowledge
Baleen: Robust Multi-Hop Reasoning at Scale via Condensed Retrieval
Omar Khattab, Christopher Potts, Matei Zaharia
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 |
|---|---|---|---|
| remove focused late interaction (FLIPR), use ColBERT retriever (ablation) | Baleen · HoVer dev | Retrieval@100: 87.4 -> 92.2 | Improved |
| remove latent hop ordering, use heuristic ordering (ablation) | Baleen · HoVer dev | Retrieval@100: 84.8 -> 92.2 | Improved |
| remove LHO, 4-hop claims (ablation) | Baleen · HoVer dev 4-hop | Retrieval@100: 65.6 -> 85.1 | Improved |
| replace condensed retrieval with re-ranking (ablation) | Baleen · HoVer dev | Retrieval@100: 92.2 -> 91.2 | Hurt |
| replace condensed retrieval with re-ranking, 4-hop claims (ablation) | Baleen · HoVer dev 4-hop | Retrieval@100: 85.1 -> 80.0 | Hurt |
| replace condensed retrieval with re-ranking, 2-hop claims (ablation) | Baleen · HoVer dev 2-hop | Retrieval@100: 97.7 -> 98.7 | Improved |
| combine condense and rerank pipelines (hybrid) (ablation) | Baleen · HoVer dev | Retrieval@100: 92.2 -> 94.5 | Improved |
| remove both FLIPR and LHO (ColBERT-Hop) (ablation) | Baleen · HoVer dev | Retrieval@100: 74.8 -> 92.2 | Improved |
| remove FLIPR and LHO, 4-hop claims (ablation) | Baleen · HoVer dev 4-hop | Retrieval@100: 47.6 -> 85.1 | Improved |
| increase number of hops from 1 to 2 (ablation) | Baleen · HoVer dev | Sentence EM: 19.7 -> 37.0 | Improved |
| increase number of hops from 3 to 4 (ablation) | Baleen · HoVer dev | Sentence EM: 38.9 -> 39.2 | No clear effect |
| use Baleen instead of ColBERT-Hop on HotPotQA | Baleen · HotPotQA fullwiki dev | P-R@20: 90.3 -> 93.3 | Improved |
Later papers that compared with this method
1 later accepted paper reports Baleen as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- MRR-FV: Unlocking Complex Fact Verification with Multi-Hop Retrieval and Reasoning (AAAI 2025 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Retriever-Free Retrieval-Augmented Reasoning via Corpus-Traversing MCTS (NeurIPS 2026 · similarity 0.88)
- MRR-FV: Unlocking Complex Fact Verification with Multi-Hop Retrieval and Reasoning (AAAI 2025 · similarity 0.88)
- Autofocus Retrieval: An Effective Pipeline for Multi-Hop Question Answering With Semi-Structured Knowledge (NeurIPS 2026 · similarity 0.88)
- Tree-of-Reasoning Question Decomposition for Complex Question Answering with Large Language Models (AAAI 2024 · similarity 0.88)
- Making Retrieval-Augmented Language Models Robust to Irrelevant Context (ICLR 2024 · similarity 0.88)
- Scaling Reasoning Hop Exposes Weaknesses: Demystifying and Improving Hop Generalization in Large Language Models (ICLR 2026 · 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.
