🔗 Causal Inference¶

🧪 ICML2026 · 19 paper notes

📌 Same area in other venues: 📷 CVPR2026 (4) · 🔬 ICLR2026 (64) · 💬 ACL2026 (7) · 🤖 AAAI2026 (7) · 🧠 NeurIPS2025 (20) · 📹 ICCV2025 (2)

🔥 Top topics: Adversarial Robustness ×2

An Odd Estimator for Shapley Values: This paper demonstrates that the Shapley value depends solely on the odd component of a set function. Based on this, it proposes OddSHAP: a method that isolates odd signals via paired sampling, screens high-order odd Fourier interactions using GBT, and performs sparse odd regression. It significantly outperforms flexible-budget Shapley estimators on mid-to-high dimensional explanation tasks.
Causal-JEPA: Learning World Models through Object-Level Latent Masking: Ours proposes C-JEPA, which extends JEPA's mask prediction from image patch-level to object-level latent representations. By using object-level masking as latent interventions, the model is forced to learn interaction-dependent dynamics. It achieves approximately a 20% gain in counterfactual reasoning over non-masked baselines and reaches comparable performance in control tasks using only 1% of tokens with over 8x planning acceleration.
Causal Modeling of Selection in Evolution: The paper argues that "selection" consists of two types: static selection (one-time filtering) and evolutionary selection (accumulation of differential reproduction over multiple generations). Existing graphical models conflate the two, leading to erroneous causal discoveries on evolutionary data. The authors define a causal graphical model that explicitly characterizes evolution and prove that its conditional independence (CI) constraints can be losslessly represented by a "clique-expanded DAG." This allows for the direct application of standard PC/GES/CDNOD algorithms, requiring only a reinterpretation of the output semantics.
Controllable Generative Sandbox for Causal Inference: This paper proposes CausalMix, a variational generative framework that jointly optimizes a type-specific multi-head decoder and a Bayesian Gaussian Mixture Model (GMM) latent prior with three independently adjustable causal "knobs" (overlap \(\alpha(X)\), CATE function \(\tau(X)\), and unobserved confounding \(\kappa(X,T)\)). While maintaining the fidelity of real-world data distributions, CausalMix allows users to design counterfactual benchmarks. Validated on real metastatic castration-resistant prostate cancer (mCRPC) patient records, CausalMix high-fidelity reproduces mixed-type tables and stably injects overlap, confounding, and heterogeneous effects as needed for controllable stress-testing of CATE estimators.
Density-Guided Robust Counterfactual Explanations on Tabular Data under Model Multiplicity: DensityFlow reformulates "generating Robust Counterfactual Explanations (RCE) under model multiplicity" as an optimal transport problem with density constraints. It uses Noise Contrastive Estimation (NCE) to train a (K+1)-way discriminator that simultaneously learns classification and class-conditional density. It then employs a Neural ODE to transport query samples along density gradients to the high-density manifold of the target class. In black-box scenarios, it aligns the surrogate only via local distillation on generated trajectories, achieving higher cross-model validity with significantly fewer queries than ensemble baselines.
ECSEL: Explainable Classification via Signomial Equation Learning: ECSEL employs "one signomial (sum of power-law terms with real exponents) per category + softmax" as a classifier. Combined with L1 sparse regularization and multi-stage optimization, it recovers 95.86% of target equations on symbolic regression benchmarks like AI Feynman with significantly lower compute than SOTA, while achieving parity with XGBoost/MLP on 11 classification datasets. All feature attributions are derived in closed-form from model parameters.
Evaluating Bivariate Causal Statements Based on Mutual Compatibility: This paper addresses scenarios where "only pairwise (bivariate) causal statements are available without ground truth." It proposes two compatibility scores that do not rely on faithfulness: comp for linear cases and incomp for graph structures. By determining whether the multivariate model formed by stitching these pairwise statements requires "anomalous extra confounding" to explain the observed covariance, the method identifies incorrect causal claims and uses it to score LLM causal outputs.
Finding Most Influential Sets: Finding the size-\(k\) subset whose removal maximizes the change in a specific estimator (Most Influential Set, MIS) originally required exhaustive search over \(\binom{n}{k}\) subsets, which is computationally intractable. This paper proves that as long as the leave-set-out effect can be expressed in a linear-fractional form, MIS selection collapses into a sequence of "top-\(k\) selection" subproblems. Utilizing Dinkelbach’s method, the approach achieves \(\mathcal{O}(n)\) per iteration with finite-step termination, providing full theoretical guarantees ranging from "exact optimality for fixed inputs" to "statistical recovery of the oracle set" within partially linear models.
Formalizing and Falsifying Causal Pathways of Rare Events: This paper formalizes "verbal causal explanations" of rare events as causal pathways—subgraphs composed of binarized events. By defining a pathway explanation score to quantify the explanatory power of "root causes + mediation pathways" relative to the target event, the authors establish a falsifiable evaluation framework for causal explanations.
From Observation to Intervention: A Causal Audit of Expert Importance in Mixture-of-Experts Models: The authors use an interventional audit of "per-token ablation" to test the implicit assumption in MoE pruning that "observational routing statistics can predict which experts are deletable." On three high-redundancy MoE models, they obtain a clean "three-model null result": none of the 60 metric-layer combinations predict the causal importance of experts after multiple-comparison correction. This suggests that existing pruning methods are effective not because metrics successfully identify "useless experts," but because redundancy in early and middle layers makes almost any selection criterion equally safe.
Investigating Memory in Model-Free RL with POPGym Arcade: This paper argues that comparing RL memory models solely by returns is unreliable. The authors construct a GPU-accelerated MDP/POMDP "twin" benchmark, POPGym Arcade, and propose four diagnostic tools: Observability Gap, Memory Bias, pixel saliency, and Recall Density. These tools reveal a "value smearing" pathology: memory models incorrectly distribute value credit across irrelevant historical observations, causing a single OOD observation to persistently contaminate policies through the recurrent state.
Outcome-Aware Spectral Feature Learning for Instrumental Variable Regression: Addressing the blind spot in Nonparametric Instrumental Variable (NPIV) regression where SpecIV learns spectral features focusing solely on the \(X-Z\) relationship while ignoring the outcome \(Y\), this paper proposes Augmented Spectral Feature Learning. By adding a regression loss of \(Y\) projected onto \(Z\) features to the contrastive loss of SpecIV, the method is equivalent to performing a truncated SVD on an "augmented operator" \(\mathcal{T}_\delta = [\mathcal{T} \mid \delta r_0]\) that incorporates \(Y\) information. This allows for causal effect recovery using extremely low-rank features even in "bad" cases where the structural function \(h_0\) is poorly aligned with the top singular functions of \(\mathcal{T}\).
Rank-Learner: Orthogonal Ranking of Treatment Effects: The authors propose Rank-Learner, the first Neyman-orthogonal two-stage treatment effect ranking learner for observational data. By replacing the indirect "estimate CATE then rank" approach with pairwise soft labels and a doubly robust correction term, it consistently outperforms T/DR-learners and non-orthogonal plug-in rankers on synthetic, semi-synthetic, and Criteo uplift real-world datasets.
Tailoring Strictly Proper Scoring Rules for Downstream Tasks: An Application to Causal Inference: This paper proposes a universal framework: by matching the local second-order curvature of the training loss \(w_\ell(p)\) with the curvature of the downstream task error \(w_{\text{task}}(p)\), one can derive a strictly proper scoring rule that is "geometrically aligned" with the downstream task. Applying this to IPW estimation of ATE yields a closed-form loss and a closed-form canonical activation function (solving a quartic equation), which consistently outperforms log-loss and covariate balancing baselines on IHDP, Jobs, Kang-Schafer, and ACIC 2017.
The (Marginal) Value of a Search Ad: An Online Causal Framework for Repeated Second-price Auctions: This paper models the true value of search advertisements as the treatment effect of "winning vs. losing." It designs an online causal learning algorithm that utilizes payment rules under binary feedback in repeated second-price auctions (SPA), achieving a minimax optimal regret of \(\widetilde\Theta(\sqrt{dT})\), which is strictly easier to learn than first-price auctions (FPA) under the same setting.
The Synthetic Web: Adversarially-Curated Mini-Internets for Diagnosing Epistemic Weaknesses of Language Agents: This paper constructs a procedurally generated "Synthetic Web" environment. By injecting a single high-credibility honeypot misinformation piece at search rank 0, it causally demonstrates that the accuracy of frontier LLM agents like GPT-5 plummets from 65% to 18% under a 1/1000 adversarial contamination. Notably, models do not increase search efforts and continue to answer with high confidence, revealing a deep-seated "positional anchoring" failure mode.
Toward Scalable and Valid Conditional Independence Testing with Spectral Representations: SpectralCIT approximates the "partial covariance operator"—a core concept in kernel methods for characterizing conditional independence—using low-dimensional spectral features learned by neural networks. It then constructs a simple HSIC-like statistic for testing. By employing a bi-level contrastive algorithm to learn the leading singular features of the operator, the authors prove that the statistic asymptotically follows a chi-squared distribution under the null hypothesis and possesses power guarantees under the alternative hypothesis. This approach effectively bridges the theoretical rigor of kernel methods with the scalability of modern representation learning.
Towards a Holistic Understanding of Selection Bias for Causal Effect Identification: This paper provides a unified "distribution-class" framework, characterizing the necessary and sufficient conditions (Condition 1) for the identifiability of the Average Treatment Effect (ATE) across the entire population under selection bias. It proves that this condition is satisfied under c-overlap propensity scores and common distributions such as polynomial exponential families, Gaussian, Laplace, Pareto, and Log-normal. Two estimators, MLE and Score Matching, are proposed with correction using a selection function \(\beta(x,y,t)\), which significantly outperform IPW and polynomial regression in synthetic and All of Us semi-synthetic experiments.
Unveiling the Structure of Do-Calculus Reasoning via Derivation Graphs: Explicitly representing all equivalent transformations of do-calculus rules through derivation graphs—revealing the structure of the causal expression space and proving that any equivalent expression is reachable within at most 4 rule applications.