AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

AdaHC: Accelerating Multi-Token Prediction with Adaptive Head Chunking with Pipeline Parallelism

Yan Wang, Chang Si, Kaiming Yang, Zhipeng Zhang, Weijian Liu, Man Yuan, Mingzhen Li, Yong Li, Weile Jia

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Distributed LLM inference systems

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 head chunking pipeline scheduling instead of naive tail placementAdaHC · Qwen3-8B (no MTP)throughput: +45.24%Improved
use AdaHC instead of Naive with 1 MTP blockAdaHC · Qwen3-30B-A3Bthroughput: +76.05%Improved
use head chunking instead of layer redistributionAdaHC · Qwen3-235B-A22B (1 MTP)throughput: +22.93%Improved
increase number of MTP blocks from 0 to 2 (ablation)AdaHC · Qwen3-8B 29-layerthroughput gain over Naive: 21.56% -> 60.08%Improved
use AdaHC instead of layer redistribution without MTPAdaHC · Qwen3-8B 29-layer, no MTPthroughputNo clear effect
vocabulary-dimension partitioning instead of sequence-dimension chunking (ablation)Vocabulary Parallel · Qwen3-8B no MTPthroughput: AdaHC +45.30% over VPHurt
move transformer layers to earlier stages (Layer Redistribute)Naive · Qwen3-8B (no MTP)throughputImproved

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