AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

CLAA: Cross-Layer Attention Aggregation for Accelerating LLM Prefill

Bradley McDanel, Steven Li, Harshit Khaitan

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs · 0 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
use GT-Ref token ranking keeping only 10% of tokensLlama-3.1-8B-Instruct · LongBench avgaverage score: 49.32 -> 49.29No clear effect
aggregate importance scores across 4 consecutive layers (CLAA) instead of a single layer (FastKV)FastKV · LongBench avg, 10% keep, Llama-3.1-8Baverage score: 46.81 -> 47.13Improved
CLAA vs FastKV difference at a fixed layer is smallCLAA · LongBench individual datasetsaverage score: +0.39No clear effect
use multi-layer aggregation instead of a single layer at the collapse layer (ablation)FastKV · LongBench six-task avg, Llama-3.2-3B, layer 11, 10% keepaccuracy: 35.05 -> 42.45Improved
use multi-layer aggregation instead of a single layer at the collapse layer (ablation)FastKV · LongBench six-task avg, Llama-3.1-8B, layer 12, 10% keepaccuracy: 41.80 -> 52.54Improved
add multi-layer aggregation (CLAA) to single-layer rankingFastKV · Needle-in-a-Haystack, Llama-3.1-8B, 40% keepaverage retrieval score: 0.906 -> 0.909Improved
add multi-layer aggregation (CLAA) to single-layer rankingFastKV · RULER 64K, Llama-3.1-8B, 40% keepaverage accuracy: 81.28 -> 82.01Improved
use a 1B speculator for lookahead (Speculative Prefill) instead of simple attention heuristicstoken ranking · LongBench avg, Llama-3.1-8B, 10% keepaverage score: 47.13 -> 41.99Hurt
push FastKV pruning layer later (19 instead of 15) (ablation)FastKV · LongBench avg, Llama-3.1-8B, 10% keepaverage score: 47.20 vs CLAA at layer 15: 47.13; TTFT 2054 ms -> 2489 msNo clear effect
use CLAA instead of semantic compression (LLMLingua, Selective Context)prompt compression · Qasper, Llama-3.1-8Bscore: 0.129 -> 0.470 (vs LLMLingua)Improved
defer KV compression for first 4 layers (m=4), costing extra KV memoryCLAA · KV cache size at 10% keepKV cache GB: 0.1 GB -> 0.3 GBHurt
use CLAA at 10% keep rate vs FullKVLlama-3.1-8B · TTFT 10k tokensTTFT: from roughly 900ms to 550msImproved

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