AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

Frustratingly Simple Retrieval Improves Challenging, Reasoning-Intensive Benchmarks

Xinxi Lyu, Michael Duan, Rulin Shao, Pang Wei Koh, Sewon Min

ICLR 2026 · Large language models › Retrieval and knowledge · approach: Retrieval-augmented reasoning · 10 citations (OpenAlex)

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 changedWhereMeasuredEffect
add CompactDS retrieval (k=10)Llama 3.1 8B Instruct · MMLU Proaccuracy: 39.8 -> 53.1Improved
add CompactDS retrieval (k=10)Llama 3.1 8B Instruct · Average over benchmarksaccuracy: 48.3 -> 55.1Improved
add exact search with GRIT after ANN retrieval (ablation)CompactDS-ANN only (k=10) · Average over benchmarksaccuracy: 53.8 -> 55.1Improved
add exact search with Contriever (same model as ANN) (ablation)CompactDS-ANN only · Average over benchmarksaccuracy: 53.8 -> 53.6No clear effect
add LM reranking (ablation)CompactDS · Average over benchmarksaccuracy: 55.1 -> 56.0Mixed
use CompactDS instead of a single Wikipedia (DPR) datastoreLlama 3.1 8B Instruct · Average over benchmarksaccuracy: 49.0 -> 54.1Improved
compress index from 456GB to 125GB (64 subquantizers) (ablation)CompactDS · Average over benchmarksaccuracy: 55.1 -> 54.1No clear effect
use oracle reranking of 100 retrieved passagesCompactDS-ANN · Average over benchmarks (6 columns)accuracy: 63.0 -> 71.2Improved
scale generator to Llama 3.3 70B Instruct and add CompactDSLlama 3.3 70B Instruct · Average over benchmarksaccuracy: 68.8 -> 71.2Improved
add CompactDS to 70B modelLlama 3.3 70B Instruct · GPQA Physicsaccuracy: 64.2 -> 62.0Hurt
add CompactDS retrievalMistral 7B Instruct · Average over benchmarksaccuracy: 37.1 -> 42.6Improved
add decontamination of search results (ablation)Search engine pipeline · MMLU Proaccuracy: 46.4 -> 42.8Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy