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
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 changed | Where | Measured | Effect |
|---|---|---|---|
| replace hard truncation with KLSC soft constraint | distribution-induced feature interpretation · 2D GMM toy, E[s]=7 | TV(q, p): 0.6198 -> 0.5095 | Improved |
| replace hard truncation with KLSC soft constraint | distribution-induced feature interpretation · 2D GMM toy, E[s]=7 | TV(q_m, q_m+eps): 0.0346 -> 0.0187 | Improved |
| replace hard truncation with KLSC soft constraint | distribution-induced feature interpretation · 2D GMM toy, E[s]=5 | boundary mass: 0.0103 -> 0.0101 | No clear effect |
| replace hard truncation with KLSC soft constraint | distribution-induced feature interpretation · 2D GMM toy, E[s]=7 | boundary mass: 0.0186 -> 0.0083 | Improved |
| use energy-guided diffusion sampling instead of retrieval or optimization with regularization | KLSC sampler · toy GMM, E[s]=7 | KL(q||p): 0.7487 (retrieval) -> 0.6574; 1.0377 (opt.+reg.) -> 0.6574 | Improved |
| use energy-guided diffusion sampling instead of retrieval or optimization with regularization | KLSC sampler · toy GMM, E[s]=5 | KL(q||p): 0.1737 (retrieval) -> 0.0577; 3.1045 (opt.+reg.) -> 0.0577 | Improved |
| use EnergyDPS instead of MACO and DiffExplainer | DINOv3 SAE features · 100 random features | QualiCLIP: 0.4228 (MACO) / 0.3259 (DiffExplainer) -> 0.5791 | Improved |
| use EnergyDPS instead of MACO | DINOv3 SAE features · 100 random features | NRQM: 6.7821 vs 4.4869 (MACO) | No clear effect |
| use EnergyDPS instead of MACO and DiffExplainer | DINOv3 SAE features · blinded study, human raters | interpretability score: 1.92 (MACO) / 1.95 (DiffExplainer) -> 2.92 | Improved |
| use EnergyDPS instead of MACO and DiffExplainer | DINOv3 SAE features · blinded study | weighted variance of descriptions: 0.2917 / 0.2643 -> 0.2505 | Improved |
| increase CLIP task-alignment guidance while intrinsic score is matched | EnergyDPS with task injection · DINOv3 SAE feature | semantic fidelity to intrinsic feature | Hurt |
| use diffusion sampling instead of optimization-based baselines | EnergyDPS · DINOv3 | compute cost | Hurt |
Closest papers by meaning
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 2024 · similarity 0.86)
- Rethinking Visual Counterfactual Explanations Through Region Constraint (ICLR 2025 · similarity 0.86)
- Data-efficient Large Vision Models through Sequential Autoregression (ICML 2024 · similarity 0.85)
- Enhancing Pre-trained Representation Classifiability can Boost its Interpretability (ICLR 2025 · similarity 0.85)
- Causal Tracing of Object Representations in Large Vision Language Models: Mechanistic Interpretability and Hallucination Mitigation (AAAI 2026 · similarity 0.85)
- Test-Time Canonicalization by Foundation Models for Robust Perception (ICML 2025 · similarity 0.85)
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.
