We propose a causal interpretation of self-attention in the Transformer neural network architecture. We interpret self-attention as a mechanism that estimates a structural equation model for a given input sequence of symbols (tokens). The structural equation model can be interpreted, in turn, as a causal structure over the input symbols under the specific context of the input sequence. Importantly, this interpretation remains valid in the presence of latent confounders. Following this interpretation, we estimate conditional independence relations between input symbols by calculating partial correlations between their corresponding representations in the deepest attention layer. This enables learning the causal structure over an input sequence using existing constraint-based algorithms. In this sense, existing pre-trained Transformers can be utilized for zero-shot causal-discovery. We demonstrate this method by providing causal explanations for the outcomes of Transformers in two tasks: sentiment classification (NLP) and recommendation.

Understanding what sparse auto-encoder (SAE) features in vision transformers truly represent is usually done by inspecting the patches where a feature's activation is highest. However, self-attention mixes information across the entire image, so an activated patch often co-occurs with--but does not cause--the feature's firing. W e propose Causal F eature Explanation (CaFE), which levarages Effective Receptive Field (ERF). W e consider each activation of an SAE feature to be a target and apply input-attribution methods to identify the image patches that causally drive that activation. Across CLIP-ViT features, ERF maps frequently diverge from naive activation maps, revealing hidden context dependencies (e.g., a "roaring face" feature that requires the co-occurrence of eyes and nose, rather than merely an open mounth.). Patch insertion tests confirm that our CaFE more effectively recovers or suppresses feature activations than activation-ranked patches. Our results show that CaFE yields more faithful and semantically precise explanations of vision-SAE features, highlighting the risk of misinterpretation when relying solely on activation location.

This paper presents a novel methodological framework for detecting and classifying latent constructs, including frames, narratives, and topics, from textual data using Open-Source Large Language Models (LLMs). The proposed hybrid approach combines automated summarization with human-in-the-loop validation to enhance the accuracy and interpretability of construct identification. By employing iterative sampling coupled with expert refinement, the framework guarantees methodological robustness and ensures conceptual precision. Applied to diverse data sets, including AI policy debates, newspaper articles on encryption, and the 20 Newsgroups data set, this approach demonstrates its versatility in systematically analyzing complex political discourses, media framing, and topic classification tasks.

The prevalence of drawing causal conclusions from observational studies has raised concerns about potential exaggeration in science communication. While some believe causal language should only apply to randomized controlled trials, others argue that rigorous methods can justify causal claims in observational studies. Ideally, causal language should align with the strength of the evidence. However, through the analysis of over 80,000 observational study abstracts using computational linguistic and regression methods, we found that causal language is more frequently used by less experienced authors, smaller research teams, male last authors, and authors from countries with higher uncertainty avoidance indices. These findings suggest that the use of causal language may be influenced by external factors such as the sociocultural backgrounds of authors and the dynamics of research collaboration. This newly identified link deepens our understanding of how such factors help shape scientific conclusions in causal inference and science communication.

Weaknesses: * This method is most suitable for variables that have a single parent in the causal DAG -- the class label. This severely restricts the class of attributes that can be modeled and manifests in the paper as experiments with simple attributes (colors in AO-CLEVr, and materials in Zappos). In fact, prior work has noted that attributes (or other compositional modifiers) manifest very differently for different objects ([36] gives the examples from prior work: "fluffy" for towels vs. dogs, "ripe" for one fruit vs. another etc.). For these attributes, and many others, the data generating process is not so straightforward -- there are edges from both attribute labels and object labels to the core features. The authors do acknowledge this limitation in L326, however it is an important weakness to consider given that _difficult_ instances in real world datasets (where both object and attribute are parents of \phi_a for example) are fairly prevalent.

We propose a causal interpretation of self-attention in the Transformer neural network architecture. We interpret self-attention as a mechanism that estimates a structural equation model for a given input sequence of symbols (tokens). The structural equation model can be interpreted, in turn, as a causal structure over the input symbols under the specific context of the input sequence. Importantly, this interpretation remains valid in the presence of latent confounders. Following this interpretation, we estimate conditional independence relations between input symbols by calculating partial correlations between their corresponding representations in the deepest attention layer.

Are generative pre-trained transformer (GPT) models only trained to predict the next token, or do they implicitly learn a world model from which a sequence is generated one token at a time? We examine this question by deriving a causal interpretation of the attention mechanism in GPT, and suggesting a causal world model that arises from this interpretation. Furthermore, we propose that GPT-models, at inference time, can be utilized for zero-shot causal structure learning for in-distribution sequences. Empirical evaluation is conducted in a controlled synthetic environment using the setup and rules of the Othello board game. A GPT, pre-trained on real-world games played with the intention of winning, is tested on synthetic data that only adheres to the game rules. We find that the GPT model tends to generate next moves that adhere to the game rules for sequences for which the attention mechanism encodes a causal structure with high confidence. In general, in cases for which the GPT model generates moves that do not adhere to the game rules, it also fails to capture any causal structure.

Intervention intuition is often used in model explanation where the intervention effect of a feature on the outcome is quantified by the difference of a model prediction when the feature value is changed from the current value to the baseline value. Such a model intervention effect of a feature is inherently association. In this paper, we will study the conditions when an intuitive model intervention effect has a causal interpretation, i.e., when it indicates whether a feature is a direct cause of the outcome. This work links the model intervention effect to the causal interpretation of a model. Such an interpretation capability is important since it indicates whether a machine learning model is trustworthy to domain experts. The conditions also reveal the limitations of using a model intervention effect for causal interpretation in an environment with unobserved features. Experiments on semi-synthetic datasets have been conducted to validate theorems and show the potential for using the model intervention effect for model interpretation.

The deconfounder was proposed as a method for estimating causal parameters in a context with multiple causes and unobserved confounding. It is based on recovery of a latent variable from the observed causes. We disentangle the causal interpretation from the statistical estimation problem and show that the deconfounder in general estimates adjusted regression target parameters. It does so by outcome regression adjusted for the recovered latent variable termed the substitute. We refer to the general algorithm, stripped of causal assumptions, as substitute adjustment. We give theoretical results to support that substitute adjustment estimates adjusted regression parameters when the regressors are conditionally independent given the latent variable. We also introduce a variant of our substitute adjustment algorithm that estimates an assumption-lean target parameter with minimal model assumptions. We then give finite sample bounds and asymptotic results supporting substitute adjustment estimation in the case where the latent variable takes values in a finite set. A simulation study illustrates finite sample properties of substitute adjustment. Our results support that when the latent variable model of the regressors hold, substitute adjustment is a viable method for adjusted regression.

In this paper we define a population parameter, ``Generalized Variable Importance Metric (GVIM)'', to measure importance of predictors for black box machine learning methods, where the importance is not represented by model-based parameter. GVIM is defined for each input variable, using the true conditional expectation function, and it measures the variable's importance in affecting a continuous or a binary response. We extend previously published results to show that the defined GVIM can be represented as a function of the Conditional Average Treatment Effect (CATE) for any kind of a predictor, which gives it a causal interpretation and further justification as an alternative to classical measures of significance that are only available in simple parametric models. Extensive set of simulations using realistically complex relationships between covariates and outcomes and number of regression techniques of varying degree of complexity show the performance of our proposed estimator of the GVIM.