AI Paper Atlas

Research map / Large language models / Retrieval and knowledge

Beyond RAG vs. Long-Context: Learning Distraction-Aware Retrieval for Efficient Knowledge Grounding

Seong-Woong Shim, Myunsoo Kim, Jae Hyeon Cho, Byung-Jun Lee

ICLR 2026 · Large language models › Retrieval and knowledge · approach: RAG robustness and evaluation

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
learn band-based retrieval of a contiguous quantile interval instead of top-k similarityLDAR · LaRA overall, 32K, open-source avgscore: 62.12 (RAG) -> 70.00Improved
retrieve a learned subset of passages instead of the full long contextopen-source LLMs · LaRA overall, 128Kscore: 43.65 -> 61.55Improved
retrieve a learned subset of passages instead of the full long contextclosed-source LLMs · LaRA overall, 32Kscore: 74.00 -> 79.42Improved
LDAR retrieves more passages than RAG on the Hallucination taskLDAR · LaRA Hallucination, 32K, open-source avgscore: 81.5 (RAG) -> 76.5Hurt
replace independent Bernoulli passage selection with band-based selection (ablation)LDAR · LaRA (training curves)performance and passage ratioImproved
add a token-count cost penalty to the training objective (ablation)LDAR · LaRA Reasoning, Qwen-2.5-7BaccuracyHurt
apply LDAR trained on LaRA to HELMET zero-shot, versus long-contextQwen-2.5-7B · HELMET HotpotQAscore: 30.0 -> 64.0Improved
apply LDAR zero-shot versus RAG on Mistral-Nemo-12B HotpotQAMistral-Nemo-12B · HELMET HotpotQAscore: 63.0 (RAG) vs 61.0No clear effect
train LDAR on bge embeddings and test with gte embeddings (ablation)LDAR · LaRA averagescore: 60.1 -> 57.7Hurt
use LDAR with transferred embedder versus RAG (ablation)LDAR · LaRA average, gte -> bgescore: 53.9 -> 55.9Improved
add Self-Route or Adaptive-k heuristics to balance RAG and long contextbaselines · LaRA overall, 32K open-sourcescore: 58.62 (LC) -> 59.45 (Self-Route), 51.70 (Adaptive-k)No clear effect

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