AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Reconstructing KV Caches with Cross-Layer Fusion for Enhanced Transformers

Hongzhan Lin, ZhiqiBai, Xinmiao Zhang, Siran Yang, Jiamang Wang, Yunlong Xu, Jiaheng Liu, Yongchi Zhao, Xiang Li, Yuchi Xu, Wenbo Su, Bo Zheng

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: KV cache eviction and compression · 6 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
fuse bottom and middle layer KV caches into top layers instead of full cacheVanilla Transformer · FineWeb-Edu valid loss, 1.5Bvalid loss: 2.241 -> 2.221Improved
fuse bottom and middle layer KV caches into top layers instead of full cacheVanilla Transformer · Downstream average, 1.5Baverage accuracy: 54.55 -> 55.82Improved
fuse KV caches (FusedKV) instead of full cacheVanilla Transformer · FineWeb-Edu valid loss, 4Bvalid loss: 2.002 -> 1.978Improved
fuse KV caches (FusedKV) instead of full cacheVanilla Transformer · Downstream average, 4Baverage accuracy: 59.07 -> 60.01Improved
use FusedKV instead of CLA at equal 50% cacheCLA · Downstream average, 650Maverage accuracy: 46.49 -> 47.45Improved
use FusedKV instead of YOCO at equal 50% cacheYOCO · Downstream average, 332Maverage accuracy: 43.08 -> 43.59Improved
use FusedKV instead of GQA at equal 50% cacheGQA · Downstream average, 1.5Baverage accuracy: 54.58 -> 55.82Improved
use FusedKV-Lite sharing instead of FusedKV (ablation)FusedKV · Downstream average, 650Maverage accuracy: 47.45 -> 48.51Improved
reverse KV assignment (K from bottom, V from middle) (ablation)FusedKV-Lite · 650M, 200B tokens, validation loss and Hellaswagvalid loss, accuracyHurt
add learnable channel-wise weights to FusedKV-Lite (ablation)FusedKV-Lite · 650M, WikiText and LAMBADA perplexityperplexityImproved
apply FusedKV on top of MLAMLA · WikiText, 650Mperplexity: 21.03 -> 20.08Improved
apply FusedKV on top of MLAMLA · downstream average, 650Maverage accuracy: 45.03 -> 44.79No 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