Research map / Probabilistic and scientific modeling / Causal discovery

Counterfactual Concept Bottleneck Models

ICLR 2025 · Probabilistic and scientific modeling › Causal discovery · approach: Counterfactual identification and bounds · 0 citations (OpenAlex)

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.

Introduces CF-CBMs, a latent-variable concept bottleneck model that jointly learns a counterfactual concept generator so one model answers "what?", "how?" and "why not?" queries, including task-driven interventions.

Problem

Concept bottleneck models answer "what?" and "how?" (concept interventions) but cannot produce concept-level counterfactuals ("why not?"), and post-hoc counterfactual generators on pixels yield hard-to-interpret changes.

Approach

Extend the CBM graphical model with a counterfactual latent z' conditioned on concepts, labels and a target class y'; train the counterfactual generator jointly with the CBM via a variational ELBO so the concept predictor yields a similar concept vector predicting an alternative class.

Key results

• CF-CBM achieves average counterfactual validity of 99.06% across five datasets, versus 96.6% for the best baseline (CBM+VCNet).
• On CUB, CF-CBM is +33 p.p. validity over BayCon and +20 p.p. over C-CHVAE; on MNIST addition it is +14 p.p. over VAE-CF.
• On modified dSprites, color confounder CaCE drops from 0.17 (post-hoc) to 0.02 (joint), while CF-CBM task/concept AUC stays close to CBM.

Measured effects reported in the paper

Surprising findings

• Jointly training the counterfactual generator with the CBM reduces reliance on confounding color concepts, shifting importance to shape concepts.
• Post-hoc counterfactual baselines applied to CBM concept space produce invalid states such as MNIST counterfactuals with zero or three active concepts.

Limitations

• Counterfactual optimality is not guaranteed compared with exact methods.
• Requires concept annotations (human or machine-generated).
• Inherits typical CBM limitations such as reasoning shortcuts and concept leakage.

Experimental setup

Datasets: dSprites MNIST addition CUB CIFAR10 SIIM Pneumothorax

Benchmarks: dSprites MNIST addition CUB CIFAR10 SIIM Pneumothorax

Metrics: ROC AUC CaCE validity proximity Delta-Sparsity IoU variability Acc Int. time

Baselines: Black Box CBM CEM CBM+BayCon CBM+C-CHVAE CBM+VAE-CF CBM+VCNet

Builds on: Concept Bottleneck Models Concept Embedding Models VCNet C-CHVAE VAE-CF

Closest papers by meaning

• Sample-efficient Learning of Concepts with Theoretical Guarantees: from Data to Concepts without Interventions (NeurIPS 2025 · similarity 0.90)
• Learning to Intervene on Concept Bottlenecks (ICML 2024 · similarity 0.89)
• Energy-Based Concept Bottleneck Models: Unifying Prediction, Concept Intervention, and Probabilistic Interpretations (ICLR 2024 · similarity 0.88)
• Causally Reliable Concept Bottleneck Models (NeurIPS 2025 · similarity 0.88)
• A theoretical design of concept sets: improving the predictability of concept bottleneck models (NeurIPS 2024 · similarity 0.87)
• From Black-box to Causal-box: Towards Building More Interpretable Models (NeurIPS 2025 · similarity 0.87)