AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Axiomatic Atlas: A Prescriptive Framework for Neural Architecture Design

Minghao Guo, Wojciech Matusik

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Mechanistic circuit interpretability

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 Atlas-guided bridges that connect the key node across the bottleneckGNN with single cross-partition edge · GNN bridge placement (m=8)accuracy (%): 51.2 -> 97.0Improved
add random bridges instead of Atlas-placed ones (ablation)GNN with bottleneck · GNN bridge placement (m=8)accuracy (%): 51.2 -> 51.8No clear effect
place bridges near the bottleneck but ignoring the key node (separator) (ablation)GNN with bottleneck · GNN bridge placement (m=8)accuracy (%): 51.2 -> 51.1No clear effect
attach bridges at the signal-carrying key node, endpoints random (ablation)GNN with bottleneck · GNN bridge placement (m=8)accuracy (%): 51.2 -> 95.8Improved
anchor quantization grid to stable percentiles (p25, p75) instead of min-max (Q-Anchor)MoE gate · MoE INT4 quantizationaccuracy (%): 11.9 -> 36.7Improved
use dynamic-k expert allocation guided by routing uncertainty instead of static k=26-layer MoE Transformer · WikiText-2validation perplexity: 228.85 -> 224.02Improved
allocate extra experts to random tokens instead of uncertain ones (ablation)dynamic-k routing · WikiText-2gap closed (%): 83% -> 22%Hurt
add orthogonal key projection (OrthoKV) removing shared mean directionattention at L=2048 · associative recallaccuracy: 0% -> 100%Improved
add OrthoKV at short contextattention at L=32 · associative recallaccuracy: 100% -> 100%No clear effect
apply gentle K-only or K+V OrthoKV projection to pretrained model (ablation)Mistral-7B-v0.1 · WikiText-103perplexity: 4.69 -> 4.75No clear effect
apply aggressive K+V projection (alpha 0.5) to pretrained model (ablation)Mistral-7B-v0.1 · WikiText-103perplexity: 4.69 -> 40.50Hurt
S2 structural check passes (no affine bypass) as a zero-cost proxyDARTS-like cells · 50 CIFAR-10 cellsaccuracy (%): 56.4 -> 70.1Improved

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