AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

FOCUS: DLLMs Know How to Tame Their Compute Bound

Kaihua Liang, Xin Tan, An Zhong, Hong Xu, Marco Canini

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Efficient language transformers · 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
add FOCUS token eviction (alpha 1.5) at confidence 0.7SDAR-8B-Chat · Math500accuracy: 54.60 -> 62.20Improved
add FOCUS token eviction at confidence 0.9SDAR-8B-Chat · Math500accuracy: 64.70 -> 64.30No clear effect
add FOCUS token eviction at confidence 0.9SDAR-8B-Chat · GSM8Kaccuracy: 89.20 -> 90.15Improved
add FOCUS token eviction vs LMDeploySDAR-8B-Chat · ShareGPT batch 256throughput (tokens/s): 979 -> 2272Improved
use block size 64 instead of 32FOCUS · ShareGPT batch 256speedup over LMDeploy: 2.32x -> 3.52xImproved
add FOCUS to LLaDA2.0-mini (MoE)LLaDA2.0-mini · ShareGPT batch 256throughput (tokens/s): 2181 -> 3369Improved
add FOCUS on LLaDA2.0 MATH at batch 64LLaDA2.0-mini · MATH batch 64throughputHurt
select tokens by top importance delta instead of random (ablation)FOCUS · GSM8K (K=2)accuracy: 40.38 -> 90.33Improved
select tokens by top importance delta instead of bottom (ablation)FOCUS · Math500 (K=2)accuracy: 4.40 -> 65.40Improved
add FOCUS eviction on top of delayed intra-block cache (ablation)SDAR-8B-Chat · MBPPaccuracy: 52.14 -> 59.73Improved
use standard delayed cache for intra-block KV (ablation)SDAR-8B-Chat · GSM8Kaccuracy: 89.20 -> 84.92Hurt
use neighbor-aware stability criterion instead of standard delayed cache (ablation)SDAR-8B-Chat · GSM8Kaccuracy: 84.92 -> 88.02Improved

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