Research map / Vision-language and multimodal / Human and embodied modeling
CoE: Chain-of-Explanation via Automatic Visual Concept Circuit Description and Polysemanticity Quantification
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Chain-of-Explanation (CoE), which uses an LVLM (GPT-4o) to automatically describe and disentangle polysemantic visual concepts of ResNet/CLIP channels, introduces a Concept Polysemanticity Entropy (CPE) metric, and generates LLM-written natural-language explanation chains along concept circuits.
Problem
Concept-based post-hoc explanations of deep vision models rely on manual, monosemantic-assumption concept labeling, which is labor-intensive and ignores polysemanticity, limiting global and local linguistic explanations.
Approach
Automatically decode channel visual concepts with an LVLM into multiple concept atoms, deduplicate them with an NLI model, filter the context-relevant atom with an LLM, quantify polysemanticity via entropy over atom distributions (CPE), and chain concept-circuit atoms into natural-language explanations.
Key results
- CoE reaches total explanation score 5.06 (GPT-4o) and 5.04 (human) versus 3.14 and 2.62 for the MILAN annotation + description baseline.
- CPE agrees with human polysemanticity judgments in 75% of 300 VC pairs.
- Averaged across scoring criteria, the paper reports a 36% average absolute improvement in explainability scores.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| apply CoE with automatic concept descriptions and contextual atom filtering instead of MILAN Annotation + Description | ImageNet explanation generation · ImageNet Validation | Total explanation score (GPT-4o): 3.14 -> 5.06 | Helped |
| apply CoE with automatic concept descriptions and contextual atom filtering instead of MILAN Annotation + Description | ImageNet explanation generation · ImageNet Validation | Total explanation score (human): 2.62 -> 5.04 | Helped |
| use automatic concept descriptions rather than MILAN annotations, without contextual filtering (ablation) | ImageNet explanation generation · ImageNet Validation | Total explanation score (GPT-4o): 3.14 -> 3.25 | Helped |
| use automatic concept descriptions rather than MILAN annotations, without contextual filtering (ablation) | ImageNet explanation generation · ImageNet Validation | Total explanation score (human): 2.62 -> 3.34 | Helped |
| add contextual concept-atom filtering (ablation) | CoE · ImageNet Validation | Total explanation score (GPT-4o): 1.81 points | Helped |
| add contextual concept-atom filtering (ablation) | CoE · ImageNet Validation | Total explanation score (human): 1.70 points | Helped |
| compare CPE polysemanticity rankings with human assessments | CPE metric · VC pairs | Agreement with human scorers: 75 % | Helped |
Surprising findings
- Polysemanticity (CPE) peaks in intermediate stages and CPE rises with model depth/complexity across ResNet50/101/152.
Limitations
- Evaluation relies largely on GPT-4o judging and a small human study of 100 samples.
- CPE human agreement is only 75%, showing moderate reliability.
Experimental setup
Datasets: ImageNet
Benchmarks: ImageNet validation (500 samples GPT-4o eval, 100 human)
Metrics: Accuracy score Completeness score User Interpretability score Total explanation score (0-6) CPE human agreement rate
Baselines: MILAN annotation + description
Builds on: CRP MILAN chain-of-thought prompting Concept Circuits
Closest papers by meaning
- Probabilistic Conceptual Explainers: Trustworthy Conceptual Explanations for Vision Foundation Models (ICML 2024 · similarity 0.88)
- Language Model as Visual Explainer (NeurIPS 2024 · similarity 0.88)
- Show and Tell: Visually Explainable Deep Neural Nets via Spatially-Aware Concept Bottleneck Models (CVPR 2025 · similarity 0.87)
- Coarse-to-Fine Concept Bottleneck Models (NeurIPS 2024 · similarity 0.87)
- FaCT: Faithful Concept Traces for Explaining Neural Network Decisions (NeurIPS 2025 · similarity 0.86)
- The Abstraction Gap in Vision-Language Causal Reasoning (ICML 2026 · similarity 0.86)
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.