AI Paper Atlas

Research map / Image generation / Autoregressive and latent generation

Residual Connections Harm Generative Representation Learning

Xiao Zhang, Ruoxi Jiang, William Gao, Rebecca Willet, Michael Maire

CVPR 2026 · Image generation › Autoregressive and latent generation · approach: Hierarchical VQ-VAE generation · 14 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
decay identity shortcut weights over encoder depth in MAEMAE ViT-B/16 · ImageNet-1Klinear probing accuracy (%): 67.8 -> 72.7Improved
decay identity shortcut weights over encoder depth in MAEMAE ViT-B/16 · ImageNet-1KKNN accuracy (%): 27.4 -> 63.9Improved
decay identity shortcut weights in MAEMAE ViT-B/16 · ImageNet-1Kfine-tuning accuracy (%): 83.6 -> 82.9Hurt
decay identity shortcuts in MAEMAE ViT-B/16 · COCO val unsupervised segmentationmIoU: 4.1 -> 10.4Improved
add encoder-decoder skip connections alongside decay (ablation)decayed MAE · ImageNet-100linear probing accuracy (%): 61.5 -> 83.6Improved
use fixed linear schedule instead of cosine schedule (ablation)decayed MAE wmin=0.6 · ImageNet-100linear probing accuracy (%): 82.8 -> 83.6Improved
use predefined fixed decay instead of learnable shortcut weights (ablation)decayed MAE · ImageNet-100linear probing accuracy (%): 79.5 -> 83.6Improved
scale only the block output branch f instead of down-weighting shortcut (ablation)MAE baseline · ImageNet-100linear probing accuracy (%): 76.5 -> 76.9No clear effect
scale both branches by constant sqrt(0.5) (ablation)MAE baseline · ImageNet-100linear probing accuracy (%): 76.5 -> 82.6Improved
apply decay to both attention and MLP rather than only one (ablation)decayed MAE · ImageNet-100linear probing accuracy (%): 79.3 (attention only) / 80.6 (MLP only) -> 83.6Improved
decay shortcuts in SiT-XL/2 diffusion with wmin=0.8SiT-XL/2 · ImageNet-1K 256x256 class-conditionalFID: 17.2 -> 11.8 at 400k iterationsImproved
add decay on top of encoder-decoder skip connections in SiT-XL/2 (ablation)SiT-XL/2 · ImageNet-1K 256x256 class-conditionalFID: 16.5 -> 11.8Improved

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