Causal Inference with the Napkin Graph

Researchers have made advancements in causal inference and emotion recognition in AI models, according to two recent studies. The 'Napkin graph' has been used to identify causal effects, while sparse autoencoders have been employed to investigate emotion-specific biases in large language models (LLMs).

A study on causal inference, submitted to arXiv on 22 Dec 2025 and revised on 25 Jun 2026^[1], introduced the 'Napkin graph', a causal structure that encapsulates features of M-bias, instrumental variables, and classical back-door and front-door settings. The method imposes a generalized independence restriction, known as a Verma constraint, rather than ordinary conditional independence restrictions. The study developed influence-function-based estimators for this functional, including doubly-robust one-step and targeted minimum loss-based estimators that remain asymptotically linear under slower-than-parametric nuisance estimation using machine learning. Another study, first submitted on 28 Apr 2026 and revised on 25 Jun 2026^[2], explored emotion recognition in LLMs. The researchers used sparse autoencoders (SAEs) to investigate emotion-specific biases and causal mechanisms of emotion inference. They found that some emotions, such as surprise and fear, rely on highly concentrated feature sets, while disgust exhibits a more distributed sparse causal organization. The study noted that LLMs are increasingly used in emotionally sensitive human-AI applications, but often perform well on some emotions while consistently struggling with others.