causal-discovery

This paper systematically evaluates 12 instruction-tuned open-weight LLMs across benchmarks to assess their reliability in classifying direct causal edges, finding they are recall-dominant, often overconfident, and that cross-prompt/model agreement improves calibration over verbalized confidence.

The paper introduces a neurosymbolic framework that uses LLMs as adaptive proposal distributions to generate causal hypotheses for modeling adverse pregnancy outcomes, combining prior knowledge with empirical data scoring to improve causal discovery.

The paper introduces SURE-Ridge, a non-iterative, closed-form method for causal discovery in equal variance linear Gaussian SEMs, achieving superior performance in small-sample and compute-limited regimes compared to existing baselines.

This paper introduces a Bayesian approach to conditional causal discovery, where the posterior over causal graphs and parameters is conditioned on user-specified causal-effect constraints (e.g., a large causal effect). They adapt rare-event estimation techniques to handle events with small posterior mass and validate the method on synthetic data and the Sachs protein dataset.

Introduces Adversarial Causal Intervention Falsification (ACIF), a sequential game where a structural causal generator proposes observational and interventional distributions while an adversarial experimentalist selects interventions to falsify it. The paper provides theoretical guarantees, including finite-sample convergence and model-selection, bridging causal generative modeling, active discovery, and experimental design.

This paper introduces GoT-CD, a Graph-of-Thoughts method for causal discovery that generates valid DAGs and shows that while structurally competitive with LLM baselines, structural fidelity alone does not guarantee fairness-faithful path-specific audits, highlighting the fragility of post-hoc path-specific fairness auditing.

AutoCause is an open-source Python framework that automates expert decisions in environmental time-series causal discovery, wrapping multiple causal-discovery methods and providing auditable, repeatable analysis. It is evaluated on 145 datasets from synthetic and real-world benchmarks.

This paper proposes a novel framework for causal discovery in multivariate time series using inverted self-attention mechanisms, introducing a causal self-attention module (CSAM) and a global causal algorithm to identify latent causal links and reduce spurious correlations.

This paper proposes a k-order relaxation of the faithfulness assumption for learning graphical Markov blankets, and introduces a proof-of-concept algorithm (kOMB) that can recover Markov blankets even under violations of faithfulness, such as parity-type relationships.

CEDAR proposes a constraint-based method for lagged causal edge discovery in sparse autoregressive time series using AR(1)-residualized distance correlation and targeted conditional-independence tests, achieving efficient edge-level interpretability with O(d²) tests after screening.

This paper studies root cause analysis on real-world telemetry data using the OpenRCA benchmark, showing that existing classical and LLM-based methods fail and proposing a Structured Multi-Agent RCA pipeline that substantially outperforms them. It further reveals through reverse reasoning that the primary bottleneck is reasoning capability rather than data access, and introduces automated rule mining to reduce reliance on manual domain knowledge.

This paper introduces DKCD, a framework that enhances causal discovery from unstructured data in high-expertise domains by integrating domain knowledge graphs with LLM-based reasoning to identify latent causal factors and improve annotation accuracy.

Proposes a finite-sample method for recovering the sparsest DAG in linear non-Gaussian acyclic models with latent confounders using higher-order cumulants, without restricting the number of latents.

NEO is a new type of world model that learns to discover reusable building blocks of explanation from raw observations without supervision or language, selected as an ICML 2026 oral presentation.

Proposes GRACE, a method that combines constraint-based skeleton with gated refinement using L0 regularization for efficient and accurate causal edge discovery in high-dimensional time series. It outperforms existing methods in F1 and speed, demonstrated on synthetic and real-world river flow data.

This survey provides a systematic review of federated causal discovery and inference, organizing methods by methodological paradigm, federation topology, and structural scope, and highlighting open challenges.

This paper argues that language model agents should assist causal discovery workflows by providing contextual support and explanations rather than generating causal conclusions, and introduces causal-learn+ platform to demonstrate this principle.

This paper proposes an interpretable causal-discovery-guided framework for deriving a Sleep Recovery Score (SRS) from multimodal polysomnography data, demonstrating up to 2.5× stronger alignment with perceived recovery than the traditional Apnea–Hypopnea Index (AHI), with potential applications in connected health.

FoundCause is an amortized causal discovery model that explicitly handles latent confounders and missing data, outperforming 15 existing methods on real-world datasets with a single forward pass.

This paper proposes LMT, a Bayesian causal discovery framework that combines LLM-extracted semantic signals from textual alarm records with timestamp-based statistical evidence to infer causal graphs in manufacturing systems.