AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Learning to Think in Physics: Breaking Shortcut Learning in Scientific Diffusion via Representation Alignment

Haozhe Jia, Pengyu Yin, Wenshuo Chen, Shaofeng Liang, Lei Wang, Bowen Tian, Xiucheng Wang, Jia Nanqian, Yutao Yue

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
add intermediate-layer PDE residual alignment (REPA-P)PIDM U-Net · Darcy flow generationdata loss: 0.0180 -> 0.0119Improved
add intermediate-layer PDE residual alignment (REPA-P)PIDM U-Net · Charge potential generationphysics residual: 0.381 -> 0.128Improved
add intermediate-layer PDE residual alignment (REPA-P)PIDM U-Net · Topology optimization IDCE%: 9.24 -> 4.17Improved
add intermediate-layer PDE residual alignment (REPA-P)PIDM U-Net · Topology optimization OODCE%: 7.93 -> 5.05Improved
add intermediate-layer PDE residual alignment (REPA-P)PIDM U-Net · Turbulent channel flowPSNR: 37.64 -> 39.95Improved
add intermediate-layer PDE residual alignment (REPA-P)PIDM DiT · Darcy flow generationdata loss: 0.0831 -> 0.0352Improved
add intermediate-layer PDE residual alignment (REPA-P)PIDM DiT · Turbulent channel flowPSNR: 38.16 -> 39.65Improved
add intermediate-layer PDE residual alignment (REPA-P)PIDM DiT · Topology optimization IDphysics residual: 1.38e-3 -> 1.31e-3Improved
align at bottleneck instead of output only (ablation)REPA-P U-Net · Darcy flow generationdata loss: 0.0162 -> 0.0119Improved
align at bottleneck instead of output only (ablation)REPA-P U-Net · Charge potential generationphysics residual: 0.188 -> 0.128Improved
align at output layer only vs bottleneck (ablation)REPA-P U-Net · Darcy flow reconstructionphysics residual: 0.0173 (bottleneck) vs 0.0142 (output)Mixed
add decoder-position alignment (ablation)REPA-P U-Net · Topology optimization IDphysics residual: 5.2e-3 -> 6.7e-3Hurt

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