AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

mHC: Manifold-Constrained Hyper-Connections

Zhenda Xie, Yixuan Wei, Huanqi Cao, Chenggang Zhao, Chengqi Deng, Jiashi Li, Damai Dai, Huazuo Gao, Mingyu Xu, Kuai Yu, Liang Zhao, Shangyan Zhou and 8 more

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Layer-wise analysis of language models · 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
constrain residual mapping to doubly stochastic manifold (mHC) instead of HCHC 27B · BBHEM: 48.9 -> 51.0Improved
constrain residual mapping (mHC) instead of HCHC 27B · DROPF1: 51.6 -> 53.9Improved
constrain residual mapping (mHC) instead of HCHC 27B · GSM8KEM: 53.2 -> 53.8Improved
constrain residual mapping (mHC) instead of HCHC 27B · MATHEM: 26.4 -> 26.0Hurt
constrain residual mapping (mHC) instead of HCHC 27B · MMLUaccuracy: 63.0 -> 63.4Improved
use mHC instead of plain residual baseline27B Baseline · 8 downstream benchmarksmixed metrics: 43.8 -> 51.0 BBHImproved
use mHC instead of HC for training stabilityHC 27B · 27B pretraininggradient norm / loss: final loss reduction of 0.021Improved
use mHC instead of HCHC · composite mapping max gainmax backward gain: nearly 3000 -> approximately 1.6Improved
add learnable H_res mapping to residual stream (ablation)residual baseline · HC component ablationabsolute loss gap: 0.0 -> -0.022Improved
add H_pre mapping on top of H_res (ablation)HC · HC component ablationabsolute loss gap: -0.022 -> -0.025Improved
add H_post mapping on top of H_res and H_pre (ablation)HC · HC component ablationabsolute loss gap: -0.025 -> -0.027Improved
scale from 3B to 27B parametersmHC · compute scalingrelative loss improvement vs baselineNo 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