Probability and conditional independence
Needed to read graphical separation, adjustment, and potential outcomes.
Causal inference
Prediction alone does not answer an intervention question. Begin by naming the two potential outcomes, the treatment assignment, the target population, and the effect to be estimated. Then ask what makes treated and control units comparable: random assignment, measured-covariate balance, a justified graph, or an explicit model assumption. This chapter follows that logic from Rubin’s finite-trial formulation and the propensity score to graphical identification, causal discovery, orthogonal estimation, heterogeneous effects, and sensitivity analysis—without treating any one estimator as a substitute for study design.
9 maintained papers · literature reviewed through 2 October 2026
Prediction estimates associations under a data distribution; causal inference asks what would change under an intervention. Identification assumptions, treatment timing, overlap, nuisance estimation, and uncertainty must be stated before an effect estimate is meaningful.
Prerequisite contract
Study design and uncertainty
Splits and metrics cannot repair confounding or positivity violations.
How the field developed
1974–1983
Potential outcomes and design-based adjustment
Rubin defines a unit-level causal effect by contrasting two potential outcomes while observing at most one; Rosenbaum and Rubin then show how a scalar treatment-assignment probability can balance observed pretreatment covariates under strong ignorability.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 01.01 Estimating causal effects of treatments in randomized and nonrandomized studies For a specified unit and two times, Rubin defines the causal effect of experimental treatment E rather than control C as the contrast between the outcome that would occur under E and the outcome that would occur under C. Only one can be observed in that trial. The paper therefore makes potential outcomes, treatment timing, and the missing comparison—not a regression coefficient—the starting point of causal reasoning. Official ETS precursor report 01.02 The central role of the propensity score in observational studies for causal effects The propensity score is the probability of receiving treatment conditional on observed pretreatment covariates. Rosenbaum and Rubin prove that it is the coarsest balancing score: at a fixed score, treated and control units have the same distribution of those observed covariates. If treatment assignment is strongly ignorable given the full covariate vector, it remains strongly ignorable given the score, enabling effect comparisons within score values. Official article record Work: Write both potential outcomes, the observed assignment, target population, overlap region, and assignment assumptions before choosing an estimator.
Exit evidence: Separate the missing potential outcome, observed-covariate balance, and unmeasured-confounding assumptions.
1995
Graphical identification
Directed acyclic graphs make causal assumptions and admissible adjustment sets inspectable.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 02.01 Causal diagrams for empirical research Pearl turns a causal directed acyclic graph into a query language for interventions. Missing arrows encode substantive exclusions; directed paths encode possible causal transmission; blocked paths encode conditional independences. Graphical criteria then decide whether an interventional quantity can be written uniquely in terms of the observed distribution. The central contribution is therefore identification before estimation: no sample size or flexible predictor can recover an effect that the assumed graph leaves non-identifiable. — Work: Draw the data-generating graph, mark observed and latent variables, then justify one adjustment set.
Exit evidence: Distinguish association, intervention, and counterfactual queries for one study.
2018
Continuous structure learning
NOTEARS turns acyclicity into a smooth equality constraint, but learned direction still depends on model and identifiability assumptions.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 03.01 DAGs with NO TEARS: Continuous Optimization for Structure Learning NOTEARS converts learning a directed acyclic graph from a combinatorial search over structures into continuous constrained optimization. Its key theorem says a weighted adjacency matrix W is acyclic exactly when the trace of the matrix exponential of the elementwise square of W equals the number of variables. A smooth equality constraint and gradient then let standard numerical optimizers search all edges jointly. Official code Work: Reproduce a synthetic-DAG experiment and stress nonlinearity, hidden confounding, and sample size separately.
Exit evidence: Report graph recovery and causal-effect error as different outcomes.
2018
Orthogonal and heterogeneous effect estimation
Double/debiased ML and causal forests use flexible nuisance models while protecting a target effect through orthogonality, honesty, or sample splitting.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 04.01 Double/debiased machine learning for treatment and structural parameters Double/debiased machine learning estimates a low-dimensional causal or structural parameter while flexible machine learners estimate high-dimensional nuisance functions. Two devices protect inference: a Neyman-orthogonal score makes the target moment locally insensitive to first-order nuisance error, and cross-fitting evaluates each observation with nuisance models trained on other folds. Together they remove regularization and overfitting biases strongly enough for root-sample-size inference under stated rate conditions. DoubleML 04.02 Estimation and Inference of Heterogeneous Treatment Effects using Random Forests Causal forests adapt random forests to estimate the conditional average treatment effect at a covariate point and to attach asymptotically valid pointwise confidence intervals. The key device is honesty: outcomes used to choose tree splits are separated from outcomes used to estimate treatment effects inside leaves. Subsampling and an infinitesimal-jackknife variance estimate then support asymptotic Gaussian inference under explicit regularity and overlap conditions. grf implementation Work: Estimate an average and conditional treatment effect with cross-fitting and overlap diagnostics.
Exit evidence: State the estimand, nuisance functions, split unit, overlap region, and interval procedure.
2026
Anytime-valid mechanistic interventions
Certified Interventional Fidelity turns patching and ablation into a causal estimand with adaptive confidence sequences under explicit score and intervention-distribution assumptions.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 05.01 Certified Interventional Fidelity: Anytime-Valid, Adaptive Evaluation of Causal Claims in Mechanistic InterpretabilityPublished in Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence, PMLR 337. Core mechanism: a declared causal estimand. The framework turns activation patching or ablation claims into an expectation over an explicit input-and-intervention distribution, then estimates it with confidence sequences that remain valid under adaptive stopping and sampling. This reading boundary matters: the certificate is statistical for the declared bounded outcome and intervention policy, not a proof that a discovered circuit is uniquely causal or valid under every intervention. — Work: Define the model-level intervention, sampled intervention distribution, score, estimand, adaptive sampling rule, and confidence sequence before running patching or ablation experiments.
Exit evidence: State the assumptions that make the sequential certificate valid and explicitly separate causal claims about model components from treatment effects in the external world.
2025
Causal adjustment with distilled 3D priors
CleanPose combines a frozen point-language teacher, a residual projector, and a proposed front-door adjustment for category-level pose estimation.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 06.01 CleanPose: Category-Level Object Pose Estimation via Causal Learning and Knowledge DistillationPublished in the ICCV 2025 main conference proceedings, pages 5990–6000. CleanPose improves category-level 6D pose estimation with two separable ideas: a causal front-door module intended to reduce spurious point-feature relations, and residual feature distillation from a frozen ULIP-2 PointBERT teacher. The zero-initialized residual adapter preserves the student's base feature while learning a teacher-aligned correction. Official implementation Work: Write the assumed causal graph, intervention, mediator, teacher information path, and residual adapter separately.
Exit evidence: Stress the front-door assumptions and compare causal, distillation, and architecture ablations on the same real and synthetic shifts.
2025–2026
Counterexamples for spatial, semantic and action grounding
Use matched distance counterexamples, cross-category part correspondence, action-absent video clips and background-consistent concealment to test different kinds of visual evidence. Preserve each study's denominator: answer probability, correspondence success, absence/presence accuracy and newly introduced ungrounded nouns are not interchangeable.
Key papers in this section
Order Primary paper · authors · year · venue/status Core idea and why to read it Official resources 07.01 Learning to Deny: Action Denial in Multimodal Large Language ModelsECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #3692 checked on 2026-09-12. Publisher version of record was not independently verified. Learning to Deny introduces UCF101-AD to test whether video-language models reject an action whose defining motion never occurs despite plausible context. Its CausalAct framework organizes persons, objects, relations, interactions and motion into an explicit verification order. The paper combines a denial benchmark with a structured-prompt and fine-tuning baseline. ECCV Spotlight program · Conference manuscript · Official project or implementation Work: Does an apparent improvement survive counterexamples that preserve superficial cues while changing the correct spatial, semantic or temporal conclusion?
Exit evidence: A small bias gap can coexist with near-chance performance. Action denial is not safety refusal, detector-based concealment is not a privacy guarantee, and benchmark correlations are not universal causal explanations.
Research branches
Potential outcomes and study design
Which unit-level contrast defines the effect, and what makes the observed treatment groups comparable?
Estimating causal effects of treatments in randomized and nonrandomized studies · The central role of the propensity score in observational studies for causal effects
Identification
Which intervention is identified from which observed distribution under which assumptions?
Causal discovery
Which graph aspects are identifiable rather than selected by an optimization bias?
DAGs with NO TEARS: Continuous Optimization for Structure Learning
Effect estimation
Is the target average, conditional, policy-specific, or transported to another population?
Double/debiased machine learning for treatment and structural parameters · Estimation and Inference of Heterogeneous Treatment Effects using Random Forests
Adaptive mechanistic intervention tests
Which internal intervention defines the estimand, and does adaptive experiment selection preserve valid uncertainty?
Causal representation transfer
Which assumptions make the adjustment causal rather than another feature-fusion regularizer?
CleanPose: Category-Level Object Pose Estimation via Causal Learning and Knowledge Distillation
Counterexamples for spatial, semantic and action grounding
Does an apparent improvement survive counterexamples that preserve superficial cues while changing the correct spatial, semantic or temporal conclusion?
Learning to Deny: Action Denial in Multimodal Large Language Models
Comparison table
| Family or stage | Scientific distinction | Main risk | Anchors |
|---|---|---|---|
| Potential-outcome definition | Defines an individual effect as a contrast between outcomes under two treatments and an average effect over specified trials or units. | Only one potential outcome is observed per unit; randomization supports inference but does not create the missing outcome or guarantee transport. | 1974 |
| Observed-covariate balancing | Compresses treatment assignment given observed pretreatment covariates into a balancing score for design or adjustment. | Balance is only for measured covariates; weak overlap, misspecified scores, and hidden causes remain. | 1983 |
| Graphical analysis | Encodes qualitative assumptions and adjustment logic. | An omitted common cause invalidates identification. | 1995 |
| Structure learning | Searches for a graph compatible with a score and assumptions. | Markov equivalence and latent variables prevent unique direction claims. | 2018 |
| Treatment-effect learning | Targets an estimand while flexibly fitting nuisance functions. | Weak overlap and data-driven subgroup search destabilize effects. | 2018 · 2018 |
| Anytime-valid mechanistic interventions | Turns mechanistic patching and ablation into an explicit model-level causal estimand with anytime-valid inference. | A valid confidence sequence cannot repair a misspecified intervention, score, sampling population, or extrapolation from model causality to real-world causality. | 2026 |
| Causal adjustment with distilled 3D priors | Couples a declared causal adjustment with residual feature transfer from a frozen multimodal 3D teacher. | Approximate front-door conditions and dataset construction may not identify the claimed real-world causal effect. | 2025 |
| Counterexamples for spatial, semantic and action grounding | Use matched distance counterexamples, cross-category part correspondence, action-absent video clips and background-consistent concealment to test different kinds of visual evidence. Preserve each study's denominator: answer probability, correspondence success, absence/presence accuracy and newly introduced ungrounded nouns are not interchangeable. | A small bias gap can coexist with near-chance performance. Action denial is not safety refusal, detector-based concealment is not a privacy guarantee, and benchmark correlations are not universal causal explanations. | 2026 |
Concept lineage
Estimating causal effects of treatments in randomized and nonrandomized studiesThe central role of the propensity score in observational studies for causal effects
The potential-outcome estimand and missing-data problem are paired with a scalar design variable that balances observed pretreatment covariates under explicit assignment assumptions.
The central role of the propensity score in observational studies for causal effectsCausal diagrams for empirical research
Propensity adjustment operationalizes balance for observed covariates; a causal graph asks whether those covariates form a valid adjustment set and exposes collider or hidden-cause risks.
Causal diagrams for empirical researchDAGs with NO TEARS: Continuous Optimization for Structure Learning
A graph changes from an explicit scientific assumption to an estimated object; identifiability does not come from optimization alone.
Causal diagrams for empirical researchDouble/debiased machine learning for treatment and structural parameters
Identification determines the estimand; orthogonal estimation then protects it from flexible nuisance-model error.
Failure modes to audit
- Treating feature importance as a causal effect.
- Conditioning on a collider or post-treatment variable.
- Ignoring interference, treatment versions, or positivity.
- Selecting subgroups and reporting intervals on the same data.
Questions for every paper
- What is the intervention and its timing?
- Which assumptions identify the target?
- Where is overlap sufficient?
- What sensitivity analysis challenges unmeasured confounding?
Complete maintained paper route
Section titled “Complete maintained paper route”9 canonical readings · 1974–2026· route reviewed through October 2, 2026.
This is the complete structured list for this chapter, ordered chronologically from earlier foundations toward current work. The explanatory ladder above highlights the shortest conceptual route; this list also includes important branches, benchmarks, and current endpoints. It is generated from the same deduplicated catalog used by the AI background registry and My Library, so a future catalog update appears here automatically.
Show all 9 readings in maintained order
Estimating causal effects of treatments in randomized and nonrandomized studies
Donald B. Rubin
Official artifacts: Official ETS precursor report
Metadata checked
The central role of the propensity score in observational studies for causal effects
Paul R. Rosenbaum and Donald B. Rubin
Official artifacts: Official article record
Metadata checked
Causal diagrams for empirical research
Judea Pearl
Metadata checked
DAGs with NO TEARS: Continuous Optimization for Structure Learning
Xun Zheng, Bryon Aragam, Pradeep K. Ravikumar, and Eric P. Xing
Official artifacts: Official code
Metadata checked
Estimation and Inference of Heterogeneous Treatment Effects using Random Forests
Stefan Wager and Susan Athey
Official artifacts: grf implementation
Metadata checked
Double/debiased machine learning for treatment and structural parameters
Victor Chernozhukov et al.
Official artifacts: DoubleML
Metadata checked
CleanPose: Category-Level Object Pose Estimation via Causal Learning and Knowledge Distillation
Xiao Lin, Yun Peng, Liuyi Wang, Xianyou Zhong, Minghao Zhu, Yi Feng, Jingwei Yang, Chengju Liu, and Qijun Chen
Published in the ICCV 2025 main conference proceedings, pages 5990–6000.
Official artifacts: Official implementation
Metadata checked
Certified Interventional Fidelity: Anytime-Valid, Adaptive Evaluation of Causal Claims in Mechanistic Interpretability
Amir Asiaee
Published in Proceedings of the 42nd Conference on Uncertainty in Artificial Intelligence, PMLR 337.
Metadata checked
Learning to Deny: Action Denial in Multimodal Large Language Models
Raiyaan Abdullah, Shehreen Azad, and Yogesh Singh Rawat
ECCV 2026 Spotlight verified in the official Vision-Language Models & Visual Foundation Models session; manuscript #3692 checked on 2026-09-12. Publisher version of record was not independently verified.
Official artifacts: ECCV Spotlight program · Conference manuscript · Official project or implementation
Metadata checked