Research map / Probabilistic and scientific modeling / Causal discovery
Causally Reliable Concept Bottleneck Models
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 Causally reliable Concept Bottleneck Models (C²BMs), which route concept predictions through an SCM whose DAG is learned from data plus LLM-refined background knowledge, with hypernetwork-parameterized linear structural equations.
Problem
Concept bottleneck models use a fixed bipartite structure that treats concepts as direct causes of the task, so they fail to reflect true causal mechanisms, limiting causal reasoning, OOD robustness, and fairness.
Approach
Structure the concept bottleneck as an SCM over a causal graph learned automatically (GES causal discovery on data, CPDAG edges oriented by a RAG-augmented LLM), with exogenous embeddings from an encoder and hypernetworks producing input-adaptive linear structural equation weights.
Key results
- C2BM matches OpaqNN and CEM task accuracy across datasets (e.g., Asia 71.4% vs 71.0% OpaqNN; CelebA 74.73% vs 74.97%) while being causally reliable.
- On Sachs, causal discovery plus LLM refinement reduces structural Hamming distance from 3.4 to 1.8 and incorrect edges from 17 to 7 versus CD alone.
- In the CelebA fairness experiment, post-intervention Causal Concept Effect of Attractive on Should be Hired drops to 0.0% for C2BM versus 8.2% to 21.9% for baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use a causal graph refined with background knowledge rather than the graph after causal discovery alone | C²BM graph construction · Synthetic causal graphs | structural Hamming distance: cMNIST: 0.2 -> 0; Asia: 0.7 -> 0.3; Sachs: 3.4 -> 1.8; Insurance: 6.4 -> 6.3; Alarm: 5.4 -> 5.0; Hailfinder: 11.0 -> 11. | Helped |
| Use a causal graph refined with background knowledge rather than the graph after causal discovery alone | C²BM graph construction · Synthetic causal graphs | incorrect edges: cMNIST: 1 -> 0; Asia: 3 -> 1; Sachs: 17 -> 7; Insurance: 19 -> 18; Alarm: 13 -> 10; Hailfinder: 22 -> 22 | Helped |
| Use a causal bottleneck instead of the strongest listed baseline | C²BM · Hailfinder | task accuracy (%): 73.4 -> 74.1 | Helped |
| Use a causal bottleneck instead of the strongest listed baseline | C²BM · cMNIST | task accuracy (%): 94.02 -> 94.18 | No clear effect |
| Use a causal bottleneck instead of the strongest listed baseline | C²BM · Sachs | task accuracy (%): 66.30 -> 65.33 | Hurt |
| Use a causal bottleneck instead of the strongest listed baseline | C²BM · CelebA | task accuracy (%): 74.97 -> 74.73 | No clear effect |
| Use a causal bottleneck instead of the strongest listed baseline | C²BM · CUB_C | task accuracy (%): 60.3 -> 58.1 | Hurt |
| Intervene on Qualified to block the Attractive -> Should be Hired path | C²BM · CelebA fairness intervention | CaCE after intervention (%): 25.1 -> 0.0 | Helped |
| Intervene on the Number concept rather than Color | C²BM · cMNIST biased ColorMNIST | Parity accuracy after ground-truth interventions: ~90% accuracy | Helped |
| Use causal relationships to propagate upstream concept interventions to downstream concepts | C²BM · Ground-truth interventions across datasets | downstream label accuracy | Helped |
Surprising findings
- The paper notes that CD alone already recovers much of the structure, and the LLM refinement mainly adds correct edges and removes spurious ones; on cMNIST the color-parity edge is removed by LLM refinement, which is the mechanism for debiasing.
Limitations
- Depends on a robust prior knowledge base, pre-trained LLMs, and well-crafted prompts; biases in knowledge or data reduce reliability.
- Scalability of state-of-the-art causal structure learning limits the approach to modest graph sizes.
- Encoder embeddings may move out of distribution, so OOD performance is not guaranteed.
Experimental setup
Datasets: Asia Sachs Insurance Alarm Hailfinder cMNIST CelebA CUB Siim-Pneumothorax
Benchmarks: Asia Sachs Insurance Alarm Hailfinder cMNIST CelebA CUB_C Siim-Pneumothorax
Metrics: task accuracy structural Hamming distance incorrect edges label accuracy under ground-truth interventions causal concept effect
Baselines: OpaqNN CBM CEM SCBM
Builds on: Concept Bottleneck Models Concept Embedding Models Greedy Equivalence Search label-free CBM Causal Concept Effect
Closest papers by meaning
- Sample-efficient Learning of Concepts with Theoretical Guarantees: from Data to Concepts without Interventions (NeurIPS 2025 · similarity 0.91)
- Learning to Intervene on Concept Bottlenecks (ICML 2024 · similarity 0.90)
- Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations (ICLR 2024 · similarity 0.90)
- Stochastic Concept Bottleneck Models (NeurIPS 2024 · similarity 0.89)
- Hyperbolic Concept Bottleneck Models (NeurIPS 2026 · similarity 0.89)
- Interpretable Concept-Based Memory Reasoning (NeurIPS 2024 · similarity 0.89)
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.