🔗 Causal Inference

💬 ACL2026 · 7 paper notes

📌 Same area in other venues: 📷 CVPR2026 (4) · 🔬 ICLR2026 (64) · 🧪 ICML2026 (19) · 🤖 AAAI2026 (7) · 🧠 NeurIPS2025 (19) · 📹 ICCV2025 (2)

🔥 Top topics: LLM ×2

Better and Worse with Scale: How Contextual Entrainment Diverges with Model Size

This paper establishes the first scaling laws for the "contextual entrainment effect," discovering that larger models are more resistant to false information in semantic contexts (negative exponent) but more prone to copying irrelevant tokens in non-semantic contexts (positive exponent), revealing opposing scaling behaviors between semantic filtering and mechanical copying functions.

ClimateCause: Complex and Implicit Causal Structures in Climate Reports

ClimateCause constructs the first expert-annotated dataset for complex and implicit causal structures in climate reports (874 causal relations), supporting nested causality, multi-event decomposition, correlation direction, and spatio-temporal context labeling. It proposes a readability metric based on causal graph semantic complexity, with LLM benchmarking revealing that causal chain reasoning remains a significant challenge.

Evaluating Counterfactual Strategic Reasoning in Large Language Models

This paper evaluates the strategic adaptation capabilities of LLMs using label perturbations, payoff perturbations, and joint counterfactual versions of the Repeated Prisoner's Dilemma and Rock-Paper-Scissors. It finds that while many models appear proficient in familiar games, they continue to apply templated strategies even after payoff structures are altered.

Function Words as Statistical Cues for Language Learning

The authors use Universal Dependencies corpora across 186 languages to demonstrate that three distributional properties—"high frequency + syntactic predictability + phrase boundary alignment"—are cross-linguistically universal. Simultaneously, they construct seven counterfactual variants of English to train GPT-2 small, proving that transformer learners perform best only when all three properties are satisfied. They identify a Goldilocks effect: function words must be both sufficiently frequent and sufficiently diverse to be both reliable and discriminative.

iTAG: Inverse Design for Natural Text Generation with Accurate Causal Graph Annotations

The iTAG framework is proposed, which utilizes a three-stage inverse design pipeline (parameterized causal graph construction → CoT-based concept assignment → structure-preserving text generation) to generate data with both extremely high causal graph annotation accuracy and text naturalness. This serves as a practical substitute for real annotated data in benchmarking text causal discovery algorithms.

Learning Invariant Modality Representation for Robust Multimodal Learning from a Causal Inference Perspective

This paper proposes CmIR (Causal Modality Invariant Representation learning), which explicitly disentangles each modality into causal invariant representations and environment-specific spurious representations based on causal inference theory. Through an elegant objective function combining invariance constraints, mutual information constraints, and reconstruction constraints, it ensures that invariant representations maintain stable predictive relationships across environments. It achieves SOTA performance in multimodal sentiment, humor, and sarcasm detection, particularly excelling in OOD and noisy scenarios.

Parallel Universes, Parallel Languages: A Comprehensive Study on LLM-based Multilingual Counterfactual Example Generation

This paper systematically investigates the multilingual counterfactual generation capabilities of LLMs across six languages. By comparing direct generation and translation-based paths, it finds that the translation path yields higher label flip rates but requires more edits. It identifies four common error patterns and validates that multilingual counterfactual data augmentation outperforms cross-lingual augmentation, particularly for low-resource languages.