AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Retrospective Sparse Attention for Efficient Long-Context Generation

Seonghwan Choi, Beomseok Kang, Dongwon Jo, Jae-Joon Kim

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression · 4 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 retrospective attention updatesQuest on Llama-3.1-8B · LongGenBench GSM8K meanaccuracy: 52.6 -> 56.5Improved
add retrospective attention updatesQuest on Llama-3.1-8B · LongGenBench MMLU meanaccuracy: 54.8 -> 55.3Improved
add retrospective attention updatesQuest on Llama-3.1-8B · LongGenBench CSQA meanaccuracy: 56.8 -> 60.3Improved
add retrospective attention updatesQuest on Llama-3.1-8B · LongGenBench CSQA n=15accuracy: 70.9 -> 68.8Hurt
add retrospective attention updatesQuest on Llama-3.1-8B · LongGenBench CSQA n=45accuracy: 46.1 -> 53.0Improved
add retrospective attention updatesQuest on Qwen2.5-14B-Instruct · LongGenBench CSQA meanaccuracy: 58.1 -> 65.5Improved
add retrospective attention updatesQuest on Qwen2.5-32B-Instruct · LongGenBench GSM8K meanaccuracy: 83.8 -> 87.8Improved
add retrospective attention updatesQuest on Qwen2.5-32B-Instruct · LongGenBench MMLU meanaccuracy: 80.0 -> 80.4No clear effect
replace top-k reuse (f=2) with retrospective update (ablation)Quest on Llama-3.1-8B · LongGenBench GSM8K meanaccuracy: 25.2 -> 56.5Improved
add top-k selection reuse with refresh every 2 steps (ablation)Quest on Llama-3.1-8B · LongGenBench GSM8K meanaccuracy: 52.6 -> 25.2Hurt
use permanent eviction (StreamingLLM) instead of QuestLlama-3.1-8B · LongGenBench GSM8K meanaccuracy: 52.6 -> 0.0Hurt
increase retrospective window w from 0 to 8 (ablation)RetroAttention · Throughput 32k batch 1batch/sec: 48.7 -> 46.2Hurt

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

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

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