AI Paper Atlas

Research map / Vision-language and multimodal / Human and embodied modeling

Escaping Low-Rank Traps: Interpretable Visual Concept Learning via Implicit Vector Quantization

Shujian Gao, Yuan Wang, Chenglong Ma, Xin Gao, Jiangtao Yan, Junzhi Ning, Cheng Tang, Changkai Ji, Huihui Xu, Wei Li, Ziyan Huang, Jiashi Lin and 7 more

ICLR 2026 · Vision-language and multimodal › Human and embodied modeling · approach: Concept bottleneck models

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 IVQ regularizer to Magnet-attention CBM (ablation)IVQ-CBM · IDRIDBMAC (%): +11.81 BMAC (45.27 -> 73.06)Improved
add IVQ regularizer (ablation)IVQ-CBM · ISIC / NCT / BUSI / Retina accuracyaccuracy (%): ISIC 80.88 -> 90.11; IDRID 57.14 -> 67.35Improved
add Magnet attention instead of [CLS] token (ablation)IVQ-CBM · medical datasetsaccuracy (%): ISIC 89.42 -> 90.11; NCT 96.23 -> 99.90Improved
use implicit quantization instead of explicit hard VQ (ablation)IVQ-CBM · ISICclassification BMAC (%): 51.47 -> 86.22Improved
use implicit quantization instead of explicit VQ (ablation)IVQ-CBM · IDRIDconcept BMAC (%): 50.85 -> 59.48Improved
use IVQ instead of Spectral Regularization (ablation)IVQ-CBM · ISICaccuracy (%): 90.21 -> 90.11No clear effect
use IVQ instead of Barlow Twins / Gram Loss / Spectral (ablation)IVQ-CBM · SIIM, BUSI, ISICBMAC (%): BUSI BMAC 91.98 -> 95.38Improved
raise feature rank via Barlow Twins or Spectral regularization beyond IVQ (ablation)IVQ-CBM · ISIC, BUSIaccuracy / BMACNo clear effect
set codebook size equal to number of text concepts (M=K) vs oversized (ablation)IVQ-CBM · medical datasets averageACC and BMACImproved
use a single shared prototype (M=1) (ablation)IVQ-CBM · medical datasets averageACC and BMACHurt
add IVQ to counter rank collapse during training (ablation)CBM visual features · ISIC, ImageNet feature rankfeature rankImproved
replace best explainable baseline with IVQ-CBMCBM · 13 benchmarksaccuracy (%): e.g. ImageNet-1K 72.29 -> 73.42; CIFAR-100 65.48 -> 67.12Improved

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