Research map / Probabilistic and scientific modeling / Causal discovery
Causal Proxy Models for Concept-based Model Explanations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| hidden-state (IIT) counterfactual objective L_HI removed (ablation) | CPM-HI · CEBaB human-created, BERT | L2 (lower better): 0.45 -> 0.79 | Helped |
| multi-task probe objective L_Multi removed (ablation) | CPM-HI · CEBaB human-created, BERT | L2 (lower better): 0.45 -> 0.47 (small) | Helped |
| random initialization instead of black-box weights (ablation) | CPM-HI · CEBaB human-created, BERT | L2 (lower better): 0.45 -> 0.81 | Hurt |
| sampled counterfactuals used instead of human-created counterfactuals | CPM-HI and CPM-IN · CEBaB BERT | L2 (lower better): 0.45 -> 0.60 (HI), 0.45 -> 0.63 (IN) | Hurt |
| CPM-HI and CPM-IN trained with approximate counterfactuals vs. best prior CEBaB explainer | CPM-HI and CPM-IN · CEBaB, all models and metrics | L2/Cosine/NormDiff: e.g. BERT L2 0.74 -> 0.45 | Helped |
| GPT-3 generated counterfactuals at inference used as explainer | GPT-3 explainer · CEBaB human-created vs sampled | L2: on-par with CPMs for human-created (0.45) but much worse for sampled (0.71 vs 0.60) | Mixed |
| random source input sampling without concept label matching at inference (ablation) | CPM-HI · CEBaB human-created, BERT | L2 (lower better): 0.45 -> 0.81 (approx., random source) | Hurt |
Closest papers by meaning
- Faithful Explanations of Black-box NLP Models Using LLM-generated Counterfactuals (ICLR 2024 · similarity 0.90)
- A New Approach to Backtracking Counterfactual Explanations: A Unified Causal Framework for Efficient Model Interpretability (ICML 2025 · similarity 0.89)
- From Black-box to Causal-box: Towards Building More Interpretable Models (NeurIPS 2025 · similarity 0.87)
- Internal Causal Mechanisms Robustly Predict Language Model Out-of-Distribution Behaviors (ICML 2025 · similarity 0.87)
- Do Models Explain Themselves? Counterfactual Simulatability of Natural Language Explanations (ICML 2024 · similarity 0.87)
- Causally Reliable Concept Bottleneck Models (NeurIPS 2025 · similarity 0.87)
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.