AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Attention with Routed-Memory for Learnable Sparse Control

QIUHAO Zeng, Jerry Huang, Peng Lu, Ruiyi Fang, Gezheng Xu, Zihao Jing, Yufei Cui, Charles Ling, Gang Niu, Boyu Wang

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression

Read the paper

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 gated soft write (fixed 1/2 sparsity) instead of FIFO eviction (ablation)ARM · Recall-intensive tasksaverage accuracy: 17.41 -> 24.42Improved
add learnable sparsity instead of fixed 1/2 retrieval (ablation)ARM · Recall-intensive tasksaverage accuracy: 24.42 -> 25.40Improved
add learnable sparsity instead of fixed 1/4 retrieval (ablation)ARM · Recall-intensive tasksaverage accuracy: 24.94 -> 25.40Improved
ARM instead of sliding window (256)Llama-3.1-8B · LongBenchaverage score: 2.88 -> 18.02Improved
ARM instead of full attentionLlama-3.1-8B · LongBenchaverage score: 17.09 -> 18.02Improved
ARM instead of full attentionLlama-3.1-8B · RULERaverage accuracy: 8.01 -> 15.12Improved
ARM instead of StreamingLLMLlama-3.1-8B · RULERaverage accuracy: 4.88 -> 15.12Improved
ARM instead of full attentionLlama-3.1-8B · LM-eval commonsenseaverage accuracy: 72.27 -> 71.89No clear effect
ARM instead of full attentionLlama-3.1-8B · WikiTextperplexity: 7.54 -> 7.86Hurt
ARM instead of full attentionLlama-3.1-8B · Recall-intensive tasksaverage accuracy: 25.31 -> 25.40No clear effect
ARM instead of full attentionLlama-3.1-8B · LongBench LCCedit similarity: 20.04 -> 11.21Hurt
ARM instead of full attentionLlama-3.1-8B · LongBench TRECaccuracy: 38.00 -> 62.50Improved

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