🔗 Causal Inference

🧪 ICML2026 · 3 paper notes

📌 Same area in other venues: 💬 ACL2026 (5) · 📷 CVPR2026 (3) · 🔬 ICLR2026 (17) · 🤖 AAAI2026 (10) · 🧠 NeurIPS2025 (21) · 📹 ICCV2025 (2)

Causal Fine-Tuning under Latent Confounded Shift

This paper proposes Causal Fine-Tuning (CFT): embedding an SCM-inspired decomposition of "high-level stable feature \(C\) + low-level confounder-sensitive feature \(\Phi\)" into standard BERT fine-tuning, and using a front-door style do-calculus adjustment formula for prediction. CFT significantly outperforms SFT/SWA/WISE and other single-domain generalization baselines under text pseudo-correlation injection attacks.

Controllable Generative Sandbox for Causal Inference

This paper proposes CausalMix: a variational generative framework that jointly optimizes a data-type-specific multi-head decoder and a Bayesian Gaussian mixture latent prior with three independently controllable causal "knobs" (overlap \(\alpha(X)\), CATE function \(\tau(X)\), unobserved confounding \(\kappa(X,T)\)). This enables users to freely design counterfactual benchmarks while maintaining fidelity to real data distributions. On real mCRPC (prostate cancer) cases, CausalMix demonstrates high-fidelity reproduction of mixed-type tabular data and stable, on-demand injection of overlap/confounding/heterogeneous effects, serving as a controllable stress test for CATE estimators.

The (Marginal) Value of a Search Ad: An Online Causal Framework for Repeated Second-price Auctions

This paper models the true value of search ads as a treatment effect between “win” and “lose” outcomes. Under binary feedback in repeated second-price auctions (SPA), it designs an online causal learning algorithm that exploits the payment rule, achieving the minimax-optimal regret \(\widetilde\Theta(\sqrt{dT})\), which is strictly easier than learning in the corresponding first-price auction setting.