AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Inside the Visual Mind: Neuroscience-Motivated Concept Circuits for Interpreting and Steering Vision Transformers

Tang Li, Yanlin Chen, Mengmeng Ma, Xi Peng

ICML 2026 · Architectures and efficiency › Attention and transformers · approach: Hybrid and interpretable ViT designs · 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
train SAE on neuroscience-motivated probing suite instead of ImageNet (ablation)SAE for ViT steering · WaterBirds worst groupworst-group accuracy (%): 63.9 -> 98.5Improved
train SAE on probing suite instead of MSCOCO (ablation)SAE for ViT steering · WaterBirds worst groupworst-group accuracy (%): 95.4 -> 98.5Improved
remove background concepts by zeroing SAE activationsCLIP-ViT-B-32 linear probe · WaterBirds worst groupworst-group accuracy (%): 50.3 -> 98.5Improved
enhance background conceptsCLIP-ViT-B-32 linear probe · WaterBirds worst groupworst-group accuracy (%): 50.3 -> 5.3Hurt
use probing-suite concept set instead of existing concept sets to interpret SAEs (ablation)SAE interpretation · Interpretation accuracy Top-30accuracy (%): 16.7 -> 54.1 (vs Google-20K)Improved
train SAE on probing suite instead of ImageNet or MSCOCO (ablation)SAE training data · Interpretation accuracy Top-30accuracy (%): 50.7 / 51.2 -> 54.1 (Ours-16K concepts)Improved
curate and deduplicate probing set across four abstraction levelsSAE training images · Concept coveragecoverage efficiency (%/1K images): 0.06 -> 1.26Improved
curate probing set vs full ImageNetSAE training images · Concept coverage averageaverage concept coverage (%): 78.9 -> 80.5Mixed
use CLIP-based concept reading instead of MLLM summarizationSAE feature auto-interpretation · ImageNet-val IoUIoU: 0.438 -> 0.432; run time 1h 40min -> 5 minNo clear effect
use BatchTopK SAE among five SAE architectures (ablation)ViT SAE · SAE benchmark on CLIP-ViT-B-32reconstruction, monosemanticity, orthogonality, dead neuronsImproved
use concept-based heatmaps instead of attribution methodconcept localization · VOC2007 QuantusPoint Game / Attribution Localization: 41.9 -> 45.0; 32.3 -> 36.0Improved

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