🔗 Causal Inference

💬 ACL2026 · 8 paper notes

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

This paper establishes the first scaling laws for "contextual entrainment," discovering that larger models better resist misinformation in semantic contexts (negative exponent) but more readily copy irrelevant tokens in non-semantic contexts (positive exponent), revealing opposing scaling behaviors of semantic filtering and mechanical copying functions.

CausalDetox: Causal Head Selection and Intervention for Language Model Detoxification

CausalDetox uses Probability of Necessity and Sufficiency (PNS) as causal criterion to precisely locate attention heads causally responsible for toxic content, applying local inference-time intervention and PNS-guided fine-tuning for detoxification, achieving up to 5.34% toxicity reduction while preserving language fluency.

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 spatiotemporal context annotation. LLM benchmarking shows causal chain reasoning remains a major challenge.

Cross-Modal Taxonomic Generalization in (Vision-) Language Models

This paper systematically studies whether LMs in VLMs can cross-modally generalize purely text-learned taxonomic knowledge (hypernym relations) to visual inputs, finding that even without any visual-language hypernym supervision, pretrained LMs can identify hypernym categories in images, but this generalization requires visual coherence among category members.

Dialectic-Med: Mitigating Diagnostic Hallucinations via Counterfactual Adversarial Multi-Agent Debate

Dialectic-Med, inspired by Popperian falsificationism, uses three-agent adversarial dialectical reasoning (proposer for diagnostic hypotheses, opponent with visual falsification module for proactively retrieving contradictory visual evidence, and mediator with weighted consensus graph), achieving SOTA on MIMIC-CXR-VQA, VQA-RAD, and PathVQA with 12.5% explanation faithfulness improvement.

Imperfectly Cooperative Human-AI Interactions: Comparing the Impacts of Human and AI Attributes in Simulated and User Studies

Through 2000 LLM simulations and a 290-person user study in a dual-framework experiment, this paper compares the impacts of human personality traits and AI design attributes in imperfectly cooperative scenarios (hiring negotiation, partially honest trading), finding that personality traits dominate in simulations while AI transparency is the key driver in real user experiments.

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

iTAG generates text with simultaneously high causal graph annotation accuracy (F1≥0.95) and naturalness (near-random detection rate) through a three-phase inverse design pipeline (parameterized causal graph construction → CoT-based concept assignment → structure-preserving text generation), serving as a practical substitute for real annotated data for benchmarking text causal discovery algorithms.

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

This paper systematically studies LLM multilingual counterfactual generation across six languages (English, Arabic, German, Spanish, Hindi, Swahili), comparing direct generation and translation paths. Translation paths yield higher label flip rates but require more edits, four common error patterns are identified, and multilingual counterfactual data augmentation outperforms cross-lingual augmentation, especially for low-resource languages.