counterfactual reasoning

While existing large language model–based social simulations can generate realistic interactions, they lack causal semantics, limiting their ability to support reliable causal inference for governance interventions. This work introduces, for the first time, a systematic integration of necessity and sufficiency–based causal concepts into social simulation, establishing a counterfactual framework tailored for policy evaluation. The framework explicitly articulates the relationship between simulator fidelity and policy relevance, offering theoretical guidance for simulator design and defining the fidelity criteria necessary to support valid policy inferences. By doing so, it advances social simulation beyond mere plausibility toward genuine decision-support capability.

This paper addresses the foundational reliance on “intervention” in causal inference by proposing a non-interventional framework for actual causality grounded in Lewis’s counterfactual semantics. Methodologically, it decomposes system states into propositional valuations and a causal basis, enabling a formal similarity relation between states—thereby defining actual causes without external interventions. Contributions include: (i) the first rigorous definition of actual causality within a purely counterfactual semantic framework; (ii) a computable model that encodes causal reasoning via propositional logic and counterfactual semantics, reducing causal verification to Quantified Boolean Formula (QBF) solving; and (iii) a complexity-theoretic result establishing that the corresponding model-checking problem is PSPACE-complete, alongside an automated verification toolchain implementation. The framework thus bridges philosophical rigor with computational tractability in actual causality analysis.

This paper addresses the lack of consensus and semantic fragmentation surrounding counterfactual definitions across causal inference (CI) and eXplainable Artificial Intelligence (XAI). We propose the first unified formal framework for counterfactuals that spans both domains. By systematically comparing their definitional logic, generation mechanisms, evaluation criteria, and application paradigms, we identify fundamental distinctions—namely, divergent semantic objectives (causal effect identification vs. model behavior explanation), contrasting constraints (structural assumptions in CI vs. feasibility or minimal perturbation in XAI), and differing practical requirements—while also uncovering complementary synergies. Our analysis clarifies core conceptual alignments and mismatches, and establishes a theoretical bridge enabling cross-domain methodological transfer: for instance, integrating structural causal models and robustness constraints from CI into XAI counterfactual generation enhances both causal fidelity and generalizability of explanations.

Rule-based intelligent environments lack mature methods for generating counterfactual explanations. Method: This paper introduces the first counterfactual explanation generation framework tailored to such environments, designed as a plug-in for existing explanation engines. It formally defines counterfactual reasoning logic under rule-based settings, constructs a generation mechanism integrating rule-based inference with causal-counterfactual hybrid logic, and proposes a context-dependent criterion for selecting explanation types. Contribution/Results: (1) It establishes a novel paradigm for actionable counterfactual explanations in rule-based intelligent environments; (2) A user study demonstrates that these explanations significantly outperform causal explanations in operational utility for problem-solving tasks, while also revealing strong contextual dependence in users’ explanation preferences.

Neural-symbolic models exhibit insufficient causal reasoning capabilities for counterfactual question answering (CQA) in videos. Method: We propose a neuro-symbolic framework jointly driven by an explicit symbolic causal graph and Answer Set Programming (ASP). Specifically: (i) we construct a symbolic causal graph to explicitly encode inter-event causal structures; (ii) we pioneer ASP-based dynamic scheduling of perception modules and language-model-driven dynamical simulation; (iii) we employ GPT-3.5/GPT-4 as promptable, interpretable counterfactual simulators. Contribution/Results: This work is the first to use ASP for logical orchestration of neural computation pipelines, endowing the model with strong causal interpretability and counterfactual controllability. Our method achieves state-of-the-art performance on CLEVRER and significantly improves counterfactual QA accuracy on CRAFT via causal-graph-guided prompting.

This work addresses the limitations of existing counterfactual explanations in abstract argumentation, which predominantly rely on but-for tests and struggle with complex causal scenarios such as preemption and overdetermination. The paper proposes an intervention-based counterfactual reasoning framework that formalizes argument acceptability conditions as structural equations and incorporates refined counterfactual criteria from the Halpern-Pearl model of actual causality. By enabling simultaneous interventions on multiple variables and integrating witness constraints to fix key argument labels, the approach achieves precise identification of intricate causal structures within abstract argumentation frameworks. This method represents the first systematic effort to handle such complexities in this domain, demonstrating significantly enhanced expressiveness and reliability compared to current approaches.

Pearl’s counterfactual semantics, grounded in structural causal models (SCMs), presupposes causal determinism and explicit response variables—limitations that hinder its applicability to unstructured probabilistic causal models. Method: We propose a novel counterfactual probability semantics for *unstructured probabilistic causal models*, requiring only real-world variables, the Markov condition, and causal completeness—without assuming determinism or designated response variables. Leveraging causal abstraction and establishing rigorous semantic equivalence, we construct, for the first time without extending SCMs, a general counterfactual interpretation framework. Contribution/Results: Our semantics reconciles the ontological disagreement between Pearl and Dawid on counterfactuals and proves semantic equivalence with both major unstructured frameworks—the potential outcomes model and the stochastic consistency model. This extends the scope of counterfactual reasoning beyond structured SCMs and provides a rigorous semantic foundation for unstructured causal modeling.

Large language models exhibit fragility in counterfactual reasoning, reflecting a lack of robust causal inference capabilities. To address this limitation, this work proposes Dual Counterfactual Consistency (DCC), a novel inference-time mechanism that enables causal evaluation and enhancement without requiring additional training or annotated data. DCC constructs dual counterfactual scenarios and integrates test-time rejection sampling to guide the model in performing causal interventions and counterfactual predictions. Extensive experiments across multiple mainstream large language models and diverse causal reasoning benchmarks demonstrate that DCC significantly improves causal reasoning performance, thereby validating its effectiveness and generalizability.

This study addresses the limited accuracy and robustness of language models (LMs) in causal reasoning—particularly counterfactual reasoning—by proposing a novel dual-agent debate framework. Two reasoning-oriented LMs, Qwen3 and DeepSeek-R1, assume complementary roles: “proponent” and “skeptic,” engaging in multi-round structured rebuttals to explicitly model competing causal hypotheses and collaboratively rectify logical flaws. To our knowledge, this is the first work to integrate reasoning-focused LMs into a causal debate mechanism. Crucially, we find that incorporating a weaker agent can effectively stimulate performance gains in the stronger agent, empirically validating the critical value of diverse perspectives in causal inference. On the CLadder benchmark, our framework achieves substantial improvements in overall accuracy (DeepSeek-R1: 78.03% → 87.45%; Qwen3: 84.16% → 89.41%), with especially pronounced gains on counterfactual reasoning tasks.

This work addresses the high computational complexity and instability inherent in counterfactual reasoning within probabilistic logic programming frameworks such as ProbLog. To overcome these challenges, the authors propose a program transformation method grounded in a weak independence assumption, which reformulates counterfactual queries into Single-World Intervention Programs (SWIPs). By structurally decomposing original clauses into observed and fixed components, the approach reduces counterfactual inference to smaller-scale marginal inference tasks. This transformation preserves correctness while substantially lowering computational overhead, and it is applicable to a broad class of structural causal models. Experimental results demonstrate that, compared to existing methods, the proposed technique reduces inference time by an average of 35%, significantly enhancing both the efficiency and reliability of counterfactual reasoning.