AI Paper Atlas

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

A Distributional View for Visual Mechanistic Interpretability: KL-Minimal Soft-Constraint Principle

Guancheng Zhou, Yisi Luo, Zhengfu He, Zhenyu Jin, Xuyang Ge, Wentao Shu, Deyu Meng, Xipeng Qiu

ICML 2026 · Vision-language and multimodal › Human and embodied modeling · approach: Language-grounded visual concept learning · 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
replace hard truncation with KLSC soft constraintdistribution-induced feature interpretation · 2D GMM toy, E[s]=7TV(q, p): 0.6198 -> 0.5095Improved
replace hard truncation with KLSC soft constraintdistribution-induced feature interpretation · 2D GMM toy, E[s]=7TV(q_m, q_m+eps): 0.0346 -> 0.0187Improved
replace hard truncation with KLSC soft constraintdistribution-induced feature interpretation · 2D GMM toy, E[s]=5boundary mass: 0.0103 -> 0.0101No clear effect
replace hard truncation with KLSC soft constraintdistribution-induced feature interpretation · 2D GMM toy, E[s]=7boundary mass: 0.0186 -> 0.0083Improved
use energy-guided diffusion sampling instead of retrieval or optimization with regularizationKLSC sampler · toy GMM, E[s]=7KL(q||p): 0.7487 (retrieval) -> 0.6574; 1.0377 (opt.+reg.) -> 0.6574Improved
use energy-guided diffusion sampling instead of retrieval or optimization with regularizationKLSC sampler · toy GMM, E[s]=5KL(q||p): 0.1737 (retrieval) -> 0.0577; 3.1045 (opt.+reg.) -> 0.0577Improved
use EnergyDPS instead of MACO and DiffExplainerDINOv3 SAE features · 100 random featuresQualiCLIP: 0.4228 (MACO) / 0.3259 (DiffExplainer) -> 0.5791Improved
use EnergyDPS instead of MACODINOv3 SAE features · 100 random featuresNRQM: 6.7821 vs 4.4869 (MACO)No clear effect
use EnergyDPS instead of MACO and DiffExplainerDINOv3 SAE features · blinded study, human ratersinterpretability score: 1.92 (MACO) / 1.95 (DiffExplainer) -> 2.92Improved
use EnergyDPS instead of MACO and DiffExplainerDINOv3 SAE features · blinded studyweighted variance of descriptions: 0.2917 / 0.2643 -> 0.2505Improved
increase CLIP task-alignment guidance while intrinsic score is matchedEnergyDPS with task injection · DINOv3 SAE featuresemantic fidelity to intrinsic featureHurt
use diffusion sampling instead of optimization-based baselinesEnergyDPS · DINOv3compute costHurt

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