mechanism identification

Core assumptions in causal discovery and inference—such as causal sufficiency, faithfulness, and the Markov condition—are inconsistently formalized and ambiguously operationalized across methodological traditions, hindering rigorous method selection under non-ideal conditions (e.g., observational constraints, limited domain knowledge). Method: We propose the first cross-framework unification framework, employing conceptual analysis, comparative modeling, and structured meta-review to systematically map how distinct paradigms model these assumptions and align them with practical inferential goals. This yields a decision guide and reusable comparison toolkit spanning the entire causal analysis lifecycle—from problem formulation to result interpretation. Contribution: Our work achieves the first cross-paradigm integration of the formal semantics and operational logic of causal assumptions. It significantly enhances the rigor and efficiency of causal methodology selection and design, particularly when background knowledge is incomplete or data are observationally constrained.

This work addresses the problem of identifying input features that are direct causes of predictions in machine learning models—critical for optimizing data collection and evaluating model interpretability. To this end, we propose the first sound and complete algorithm for direct cause discovery, grounded in structural causal models (SCMs). Our method integrates conditional independence testing, constraint satisfaction solving, and a novel causal independence rule, substantially improving search efficiency. We provide formal theoretical guarantees establishing both correctness (soundness) and completeness. Empirical evaluation across diverse benchmark datasets demonstrates that our algorithm achieves significantly higher accuracy and computational efficiency in identifying direct causes compared to state-of-the-art alternatives. The results establish our approach as a robust, theoretically grounded tool for behavioral analysis of predictive models and causal-aware data engineering.

This work addresses the problem of inferring interpretable causal graphs from data under prescribed constraints on causal effects. The authors formulate constrained causal discovery as a conditional Bayesian inference task, aiming to estimate the posterior distribution over causal graphs and parameters given specific causal effect conditions—such as requiring large effects. To tackle the computational challenges posed by low-probability constraint regions, they innovatively integrate rare-event estimation techniques with particle filtering and graph structure learning. Experimental results demonstrate that the proposed method achieves both accuracy and scalability on synthetic data, and successfully generates biologically meaningful pathway-level causal summaries on the Sachs protein signaling dataset, substantially enhancing the interpretability and practical utility of causal discovery.

Modeling dynamic, state-dependent causal relationships in neuroscience remains challenging, as existing methods typically assume static and linear causal structures—limitations that hinder accurate characterization of time-varying interactions in complex systems like the brain. Method: We propose the Conditional Weighted Static Graph Overlay (CW-SGO) framework, the first to jointly integrate generative factor models, dynamic graph neural networks, and nonlinear causal inference for end-to-end learning of nonlinear, time-varying causal graphs. CW-SGO explicitly models how latent brain states modulate causal connectivity, thereby relaxing restrictive assumptions of stationarity and linearity. Results: Evaluated on synthetic benchmarks and real human fMRI data, CW-SGO achieves average F1-score improvements of 22–28% over baselines, with gains exceeding 60% in specific tasks. It successfully uncovers task- and rest-state-specific functional connectivity patterns, demonstrating strong interpretability and neuroscientific validity.

Traditional causal discovery methods in dynamic systems are often constrained by assumptions of acyclicity or stationarity, rendering them ill-suited for real-world complexities such as feedback loops, cyclic interactions, and nonstationarity. This work proposes a hybrid causal discovery framework that integrates physical knowledge with data-driven learning by embedding known physical mechanisms as inductive biases into a stochastic differential equation (SDE) model: the drift term encodes established ordinary differential equation (ODE) dynamics, while the diffusion term captures unknown causal couplings. By systematically incorporating partial physical priors into dynamic causal discovery for the first time, the approach overcomes conventional limitations and substantially enhances the identifiability and robustness of the inferred causal graph. Coupled with sparsity-inducing maximum likelihood estimation, the proposed algorithm outperforms state-of-the-art purely data-driven baselines across multiple dynamic system benchmarks, yielding more accurate, stable, and physically consistent causal structures.

This work addresses a fundamental limitation in traditional behavioral measurement, which often relies on passive observation under static or weakly controlled conditions and struggles to disentangle distinct internal mechanisms that produce similar overt behaviors. Treating human behavior as the observable output of a dynamic system, this study introduces— for the first time—the principles of system identification into behavioral science. It proposes a closed-loop experimental framework based on structured perturbations: precise, programmable disturbances are delivered via immersive environments while multimodal behavioral trajectories are simultaneously recorded. These data are integrated with dynamic computational models to enable mechanism-driven, real-time inference. By synergistically combining psychometrics, experimental design, and generative modeling, the approach advances behavioral science from descriptive analysis toward an identifiable, reproducible paradigm centered on generative mechanisms, substantially enhancing both theoretical rigor and causal interpretability in behavioral inference.

This work addresses the challenge of formalizing multiscale causal relationships in complex systems by proposing a concise discrete hierarchical causal modeling framework. The framework introduces causal classes to abstract cross-level causal influences and integrates aggregation operators with discrete event-time mappings to characterize how high-level actors constrain, select, and organize the behaviors of low-level agents. The resulting formalism comprises three core components—causal classes, aggregation mechanisms, and temporal mappings—providing a unified and computationally tractable foundation for hierarchical causal analysis in complex systems.

This work addresses the challenge of causal discovery in dynamic systems where delayed or overlapping causal effects render traditional observational methods ineffective. Focusing on chain-reaction systems characterized by cascading activations, the authors propose a causal identification strategy based on blocking interventions: by selectively preventing component activation, the true causal structure can be uniquely determined with only a small number of targeted interventions. Theoretical analysis demonstrates that the proposed method achieves exponential error decay and logarithmic sample complexity under finite-sample conditions. Empirical evaluations on both synthetic models and diverse chain-reaction environments confirm its efficacy, substantially outperforming purely observational heuristic approaches.

This work addresses the challenge of uncovering causal mechanisms underlying macroscopic emergent phenomena in large language model (LLM) agent systems, which arise from complex microscopic interactions and nonlinear dynamics and are thus difficult to interpret. To this end, the paper introduces CAMO, a novel framework that, for the first time, automatically discovers interpretable causal chains linking micro-level agent behaviors to macro-level emergence. CAMO formalizes mechanistic hypotheses as computable factors, learns the Markov boundary of a target variable Y along with its minimal upstream explanatory subgraph, and employs in-simulator counterfactual probes to directionally disambiguate causal edges and refine hypotheses. Experiments across four canonical emergence scenarios demonstrate that CAMO efficiently identifies concise, interpretable, and intervention-supportive causal mechanisms, offering a new paradigm for causal understanding and control of complex agent systems.

This work addresses the lack of a unified validation framework for evaluating whether high-level causal abstractions faithfully reflect underlying mechanisms. The authors construct a benchmark encompassing ten classes of complex systems—spanning discrete/continuous and static/dynamic types—and systematically assess over thirty metrics under a common causal abstraction framework to distinguish valid from invalid abstractions. They propose a novel continuous measure, Causal Abstraction Error (CAE), which passes discriminative tests across all systems and converges with only 30 interventions. Additionally, they introduce a fidelity test for unmapped variables that integrates observational, functional, information-theoretic, and causal criteria. Experiments demonstrate that causal metrics constrained solely by faithfulness reliably discriminate abstraction validity, with CAE exhibiting both superior performance and computational efficiency.