🔗 Causal Inference¶

🤖 AAAI2026 · 10 paper notes

CaDyT: Causal Structure Learning for Dynamical Systems with Theoretical Score Analysis: This paper proposes CaDyT, which combines Gaussian process-based continuous-time dynamics modeling (via Adams-Bashforth integrators for exact inference) with the Minimum Description Length (MDL) principle for structure search. The method simultaneously addresses irregular sampling and causal structure identification, substantially outperforming all baselines on double-mass spring, diamond graph, and Rössler oscillator benchmarks (AUPRC 0.79 vs. runner-up 0.39).
Causally-Grounded Dual-Path Attention Intervention for Object Hallucination Mitigation in LVLMs: This paper proposes Owl, a framework that models visual and textual attention as mediating variables within a structural causal model, introduces the VTACR metric to quantify cross-modal attention imbalance, and designs VTACR-guided adaptive attention modulation combined with a dual-path contrastive decoding strategy, achieving state-of-the-art hallucination mitigation on POPE and CHAIR benchmarks.
From Theory of Mind to Theory of Environment: Counterfactual Simulation of Latent Environmental Dynamics: This paper proposes the concept of "Theory of Environment" (ToE), arguing that humans may infer latent environmental dynamics through computational mechanisms shared with Theory of Mind (ToM), thereby expanding the dimensionality of motor exploration and facilitating behavioral innovation.
Hallucinate Less by Thinking More: Aspect-Based Causal Abstention for Large Language Models: This paper proposes ABCA (Aspect-Based Causal Abstention), a pre-generation abstention framework that employs dual-agent debate to identify "aspect variables" (e.g., discipline, legal context, temporal frame) for activating distinct knowledge branches within LLMs. It applies the AIPW doubly robust estimator to compute causal effects and uses Centroid Angular Deviation (CAD) to detect knowledge conflicts (Type-1) or knowledge insufficiency (Type-2), achieving 91.4% accuracy on TruthfulQA and 96.4% unanswerable question identification rate—far surpassing the baseline of 44%.
I-CAM-UV: Integrating Causal Graphs over Non-Identical Variable Sets Using Causal Additive Models with Unobserved Variables: This paper proposes I-CAM-UV, a method that enumerates consistent DAGs satisfying structural constraints derived from multiple CAM-UV outputs over non-identical variable sets, recovering causal relations lost due to unobserved variables, and introduces an optimal-first search algorithm exploiting cost monotonicity for efficient combinatorial search.
KTCF: Actionable Recourse in Knowledge Tracing via Counterfactual Explanations for Education: This paper proposes KTCF, a counterfactual explanation generation method for Knowledge Tracing (KT) that leverages inter-concept relationships to produce sparse and actionable counterfactuals, subsequently post-processed into sequentially ordered instructional recommendations. KTCF comprehensively outperforms baseline methods across validity, sparsity, and actionability metrics.
Learning Subgroups with Maximum Treatment Effects without Causal Heuristics: Under the SCM framework, the paper proves that the subgroup with maximum treatment effect must exhibit homogeneous pointwise effects (Theorem 1); under the partition model assumption, it proves that optimal subgroup discovery reduces to standard supervised learning (Theorem 2), achievable via CART with the Gini index. On 77 ACIC-2016 semi-synthetic datasets, the proposed method achieves a mean treatment effect of 10.54 (vs. 7.84 for the runner-up), ranking first on 51.9% of datasets.
MUG: Multi-agent Undercover Gaming — Hallucination Removal via Counterfactual Test for Multimodal Reasoning: MUG reframes Multi-Agent Debate (MAD) as a "Who's Undercover" social reasoning game — by introducing information asymmetry through counterfactual image editing (modifying the reference image), one agent is assigned the edited image \(I^-\) as the "undercover," while other agents hold the original image \(I^+\) and identify the undercover (i.e., the hallucination source) via reasoning and voting. On HallusionBench, Qwen2.5VL-7B improves from 46.4% to 53.8%.
Skill Path: Unveiling Language Skills from Circuit Graphs: This paper proposes the concept of Skill Path and a three-step framework (Decomposition–Pruning–Causal Mediation) to extract linear paths corresponding to specific language skills from circuit graphs, and quantitatively validates two core conjectures: Stratification and Inclusiveness of skills.
Sparse Additive Model Pruning for Order-Based Causal Structure Learning: This paper proposes SARTRE, a framework that employs randomized tree embeddings and group-sparse regression to learn sparse additive models, replacing the hypothesis-testing-based redundant edge pruning in CAM-pruning for order-based causal structure learning. SARTRE achieves significant speedups without sacrificing accuracy.