Skip to content

๐Ÿ”— Causal Inference

๐Ÿง  NeurIPS2026 ยท 41 paper notes

๐Ÿ“Œ Same area in other venues: ๐ŸŽž๏ธ ECCV2026 (3) ยท ๐Ÿ“ท CVPR2026 (4) ยท ๐Ÿ”ฌ ICLR2026 (64) ยท ๐Ÿ’ฌ ACL2026 (7) ยท ๐Ÿงช ICML2026 (19) ยท ๐Ÿค– AAAI2026 (7)

๐Ÿ”ฅ Top topics: Adversarial Robustness ร—3

B-CALM: Bias-Limited Bayesian Borrowing for RCT-Anchored Treatment Effects under Covariate Mismatch

Addressing the challenge where randomized controlled trials (RCTs) lack sample size for conditional average treatment effect (CATE) heterogeneity and observational studies (OS) are confounded with partially overlapping covariates, B-CALM proposes a Bayesian calibrated alignment and borrowing framework that decouples baseline and comparative bias functions, establishes a closed-form information ceiling and saturating effective sample size on observational contrast borrowing, and maintains nominal credible interval coverage without negative transfer.

Beyond the Training Distribution: Evaluating Predictions Under Distribution Shift and Selection Bias

Addressing the compounded challenges of covariate shift and historical decision-induced selective labels in pre-deployment model evaluation, this paper establishes a joint identification framework and derives an efficient influence function to construct an asymptotically normal and semiparametrically efficient double machine learning estimator.

Budgeted Multi-Source Counterfactual Annotation for Off-Policy Evaluation

Addressing limited behavior-policy coverage in contextual-bandit off-policy evaluation, this paper formulates a budget-constrained multi-source counterfactual annotation allocation problem, establishes a first-annotation threshold and four value regimes, and designs a majorization-minimization algorithm that monotonically minimizes estimator variance, reducing evaluation MSE by up to 20.68%.

Causal Abstractions, Categorically Unified

A unified algebraic framework for causal abstractions built upon Markov categories and deterministic natural transformations, rigorously subsuming and generalizing deterministic, probabilistic, non-aligned, and confounded graphical abstractions.

Causal Discovery Under Hard Selection Bias: A New Robust Score-Matching Approach

Addressing observational causal discovery under arbitrary, unknown hard selection bias (truncation), this paper proves that log-density scores and their Hessians remain strictly invariant in the interior of the selection support, and introduces CHessโ€”a robust score-based approach using conditional Hessian rank testing that identifies full causal DAGs across linear and non-linear Additive Noise Models without any selection process knowledge.

Causal Discovery via Transformed Low-Rank Quantile Surfaces

Proposes Low-Rank Quantile Surfaces (LRQS), establishing generic identifiability for bivariate causal discovery under unknown monotone observation distortions and multirank latent shape variations, paired with an alternating truncated SVD and isotonic regression scoring algorithm.

Causal Inference for Sequential Settings under Interference and Latent Confounding

Addressing observational sequential settings plagued by spatial treatment interference, temporal autocorrelation, and low-rank latent confounding, this paper proposes a differentiable framework combining sequential Maximum Pseudo-Likelihood Estimation (MPLE) with low-rank alternating minimization, establishing finite-sample non-asymptotic parameter consistency and generalized treatment effect identification from a single high-dimensional trajectory.

Causal Learning with the Invariance Principle

This paper proves that under acyclicity and bijectivity assumptions, the invariance principle requires only two auxiliary environments (natural experiments) to identify the causal graph and functional mechanisms for arbitrary nonlinear SCMs, guaranteeing identifiability at the counterfactual level.

Causal Multi-Task Demand Learning

To tackle endogeneity in multi-task demand learning where posted prices are systematically correlated with unobserved store-level demand drivers, this paper proposes Decision-Conditioned Masked-Outcome Meta-Learning (DCMOML), which conditions on all endogenous prices while enforcing dual outcome masking and randomized query supervision over locally exogenous price pairs to consistently identify the conditional mean causal parameters.

Causal Representation Learning for Generalisable Recommendation

This paper introduces an information-theoretic causal representation learning objective for recommendation under policy-induced distribution shift, penalising mutual information between item representations and user confounders to discard non-causal leakage and delivering significant online gains on Spotify without any inference-time overhead.

Causal Survival Forests with Negative Controls

Negative Control Causal Survival Forests (NC-CSF) integrates proximal causal inference into causal survival forests, using Neyman-orthogonalized residuals to deliver robust, nonparametric heterogeneous treatment effect estimation in observational survival studies with both right censoring and unmeasured confounding.

CausalCompass: Evaluating the Robustness of Time-Series Causal Discovery in Misspecified Scenarios

Introduces CausalCompass, a comprehensive and extensible benchmark evaluating 11 representative time-series causal discovery algorithms across 8 assumption-violation scenarios, demonstrating through 110,000+ experimental runs that deep learning methods consistently provide superior robustness and highlighting the counterintuitive benefits of standardization on continuous-optimization methods.

CausalMix: Data Mixture as Causal Inference for Language Model Training

CausalMix reformulates supervised fine-tuning (SFT) data mixture optimization as a state-conditioned causal marginal return estimation problem, isolating data pool confounding biases via Double Machine Learning and causal forests to extrapolate optimal mixture ratios across data scales and model architectures without expensive proxy retraining.

Coarsening Linear Non-Gaussian Causal Models with Cycles

This paper extends causal coarsening theory to linear non-Gaussian structural causal models with feedback cycles, demonstrating that the strongly connected component (SCC) partition forms the structural floor of the valid DAG-coarsening lattice and that the resulting condensation DAG is uniquely identifiable from observational data in worst-case cubic time.

Conditional Counterfactual Mean Embeddings: Doubly Robust Estimation and Learning Rates

This paper introduces Conditional Counterfactual Mean Embeddings (CCME) to represent conditional counterfactual distributions within reproducing kernel Hilbert spaces, establishing a doubly robust two-stage meta-estimator with three practical implementations and proving finite-sample minimax optimal convergence rates under Sobolev smoothness.

Conservative Continuous-Time Treatment Optimization

Addressing offline treatment optimization from irregularly sampled longitudinal patient trajectories, this paper models patient dynamics via controlled stochastic differential equations (Controlled SDEs) and constructs a conservative path-space signature-kernel maximum mean discrepancy (Signature-Kernel MMD) regularizer to strictly penalize out-of-support treatment plans and avoid model exploitation.

Controllable Generative Sandbox for Causal Inference

CausalMix unifies empirical tabular data fidelity with explicit causal controllability through a variational generative sandbox coupling Bayesian Gaussian mixture priors with variable-specific decoders, log-density ratio overlap regularization, and robust CATE / confounding constraints.

Controlling for Omitted Variable Bias in Deep Neural Networks

Formalizing shortcut learning in deep neural networks as classical omitted variable bias, this paper refits the final linear layer of a frozen pretrained backbone into an additive model with ridge penalization and cross-fitting to achieve unbiased direct effect estimation and marginalized unconfounded predictions.

Covariate-Adjusted Deep Causal Learning for Heterogeneous Panel Data Models

This paper introduces Covariate-Adjusted DEep CAusal Learning (CoDEAL), a framework that purges complex nonlinear time-varying covariate effects via a feed-forward deep neural network and extracts latent nonlinear factor dynamics via a multi-output autoencoder, achieving accurate counterfactual imputation and unit-specific heterogeneous ATT estimation under staggered adoption designs.

DARTS: Targeting Prognostic Covariates in Budget-Constrained Sequential Experiments

Addressing sequential randomized controlled trials under high-dimensional costly covariates, this paper introduces DARTS, a framework that employs combinatorial semi-bandits with knapsacks via Thompson sampling to adaptively select prognostic covariates for rerandomization and regression adjustment, closing the efficiency gap to oracle designs while strictly guaranteeing design-based unconfoundedness and asymptotic conservative coverage.

Data-Driven Covariate Selection for Nonparametric and Cycle-Agnostic Causal Effect Estimation

This paper proves that local, data-driven covariate selection and causal effect estimation based on conditional independence within Markov blankets remain sound and complete in simple structural causal models with feedback loops, establishing a unified cycle-agnostic framework.

Debiased Counterfactual Generation via Flow Matching from Observations

This paper reveals the structural isomorphism between observational conditional distributions and counterfactual outcome distributions in terms of support, tail behavior, and confounder-invariant features, proposing a flow-matching framework that learns "deconfounding flows (DecFM)" starting directly from observational conditional distributions via a novel efficient influence function (EIF) debiased gradient estimator and a minibatch entropic optimal transport (EOT) extension, substantially outperforming existing debiased generative methods across causal benchmarks and high-dimensional image attribute rebalancing tasks.

Generalized Bayes for Causal Inference

To overcome the vulnerabilities of high-dimensional nuisance likelihood modeling in Bayesian causal inference, this paper proposes a generalized Bayes framework with identification-driven Neyman-orthogonal losses, placing priors directly on causal estimands to deliver robust, frequentist-calibrated uncertainty quantification.

Graph Distance Based on Cause-Effect Estimands with Latents

To address the fundamental challenge of evaluating causal graphs under unobserved confounding, this paper introduces the Fixing Identification Distance (FID)โ€”the first cause-effect estimand-based distance measure for acyclic directed mixed graphs (ADMGs)โ€”using identification via fixing and a canonical symbolic verifier.

Identifying Structural Biases from Causal Mechanism Shifts

This paper demonstrates that hidden confounding and selection bias induce distinct, topologically asymmetric dependent mechanism shifts across multiple contexts, and introduces StruBI, a non-parametric algorithm using adjusted mutual information and ancestral coverage to accurately identify structural biases and recover affected variable subsets.

Impossibility of Distribution-Free Predictive Inference for Individual Treatment Effects

This paper establishes finite-sample and asymptotic impossibility theorems demonstrating that any distribution-free prediction set achieving valid coverage for individual treatment effects (ITEs) in the presence of continuous covariates must be trivial with infinite expected length, uncovering the fundamental connection between counterfactual missing data and the hardness of conditional independence testing.

Integrating Background Knowledge for Scalable Causal Discovery

Addressing the bottleneck where background knowledge is confined to post-processing and fails to reduce computational complexity, this paper proposes integrating direct causal background knowledge during the causal discovery process via separating set pruning, skipping known adjacencies, and delaying gap evaluations, reducing independence test requirements by orders of magnitude with theoretical soundness and completeness guarantees.

Just Ramp-Up: Debiasing Regression-based Estimator for A/B Tests under Network Interference

Addressing the extreme extrapolation challenge in estimating Global Average Treatment Effects (GATE) under network interference, this paper shows that pooling data from multi-stage ramp-up experiments with diverse treatment proportions enriches micro-level network exposure variation, effectively debiasing both linear and graph neural network estimators at zero additional experimental cost.

learning causal orderings for in-context tabular prediction

TABORDER proposes an in-context tabular prediction framework that simultaneously infers causal topological orders via an unsupervised likelihood-based objective and enforces causal order-constrained column attention alongside missingness-aware variance modeling, achieving robust predictions under mechanism interventions.

Learning Deployable Causal Action Geometry under Temporal Non-Stationarity

To address the fundamental flaw where temporal non-stationarity in observational time series causes pooled predictive models to reverse action rankings, this paper introduces deployable causal action geometry that anchors away baseline shifts and profiles out positive amplitude variation, coupled with a two-stage honest orthogonal estimator that achieved a +1.16% GMV lift and +1.70% buy counts lift in a real-world Taobao pricing A/B test.

MapPFN: Learning Causal Perturbation Maps in Context

MapPFN is the first prior-data fitted network for single-cell causal perturbation prediction, meta-learning mappings between observational and interventional distributions on a synthetic biological prior to adapt to unseen biological contexts and arbitrary gene sets at inference time purely through in-context learning.

Offloading Score: Measuring AI Reliance through Counterfactual Workflows

By reconstructing human-only counterfactual workflows to estimate steps saved through AI assistance, this work introduces Offloading Score, a process-oriented reliance metric that quantifies cognitive offloading and empirically benchmarks overreliance boundaries.

Orthogonal Uplift Learning with Permutation-Invariant Representations for Combinatorial Treatments

Addressing the issue where combinatorial policy interventions in industrial decision-making are treated as categorical IDsโ€”inducing estimation bias and poor sample sharingโ€”this paper introduces POUL, which embeds policies via a permutation-invariant aggregation over context-action mixtures and integrates them into a factorized orthogonalized uplift model with second-order robustness against nuisance errors.

PAIR-CI: Calibrated Conditional Independence Testing for Causal Discovery with Incomplete Data

To eliminate severe false-positive miscalibration caused by the "impute first, test second" paradigm in causal discovery with missing values, PAIR-CI introduces a nonparametric conditional independence test integrating multiple imputation with paired cross-validated permutations to cancel residual imputation bias, accompanied by the first formal unification of cross-validation and Rubin's variance estimation.

Prediction-Intervention Games and Invariant Sets

This paper formalizes strategic prediction under structural causal models as a Stackelberg "prediction-intervention game", showing that invariant predictors based on the stable blanket weakly dominate causal parents across common follower utilities and attain minimax worst-case risk optimality under graphical and strong-intervention conditions.

Querying Counterfactuals on Tissue Graphs with Supervised Disentanglement

This paper formalizes tissue graph counterfactuals via edge rewiring and neighbor node perturbations and introduces Cellina, a dual-encoder graph VAE with supervised disentanglement that accurately predicts single-cell transcriptional shifts under altered spatial microenvironments.

Reducing Credit Assignment Variance via Counterfactual Reasoning Paths

Introduces Implicit Behavior Policy Optimization (IBPO), a framework that leverages counterfactual comparisons and local self-corrections across multiple reasoning paths to construct implicit process-level advantages and token-level gradient masks, substantially reducing credit assignment variance under sparse terminal rewards without requiring step-level annotations or value models.

SCDM: Scalable Causal Discovery in Nonlinear Temporal Systems with Meta-Learning

SCDM reframes nonlinear temporal causal discovery as shared-parameter episodic meta-learning over target-variable prediction tasks, coupling task subsampling with a shared causal-strength matrix to achieve accurate directed graph recovery across thousands of variables with remarkably low runtime and memory.

Structure-agnostic Causal Representation Learning

Addressing the severe predictive signal loss caused by misspecified fixed invariance constraints when causal structures are unknown, this paper proposes SaCRLโ€”a framework that jointly identifies the underlying causal structure and learns invariant representations via HSIC violation metrics and adaptive SoftMin weighting, achieving state-of-the-art out-of-distribution generalization across synthetic benchmarks and DomainBed.

Towards Identifiable Latent Additive Noise Models

This paper overcomes the restrictive functional assumptions of existing causal representation learning by showing that environment-induced changes in causal influences guarantee component-wise identifiability for latent additive noise and post-nonlinear models, accompanied by an ELBO-based variational inference algorithm evaluated on synthetic, fMRI, and human motion capture data.

Where Root Cause Analysis Fails: A Retrieval-Reranking Decomposition

The paper separates root cause ranking accuracy into whether the cause enters the candidate set and whether it ranks highly once retrieved, audits six suites from four datasets, and uses multi-signal retrieval plus single-call LLM reranking to show why the two failures require different remedies rather than simply more causal structure or reasoning capacity.