AnTenA: Actionable and Explainable Tensor Analysis System with Large Language Models

Researchers have detailed AnTenA, a system that uses large language models to produce human-readable explanations of hidden patterns extracted from complex, multi-aspect data through tensor decomposition [1]. The system, formally named the Actionable and Explainable Tensor Analysis System, was described in a paper submitted on 27 Jun 2026 [1]. It is designed to address a persistent challenge in data analysis: the metadata or labels needed to interpret latent patterns are often inaccurate, inconsistent, or entirely unavailable [2]. AnTenA bypasses this by leveraging the knowledge embedded within large language models (LLMs) to generate explanations in plain language [1]. The system operates in two stages. The first module performs tensor analysis, using techniques such as CANDECOMP/PARAFAC decomposition to uncover latent components that capture hidden relationships between different modes of data [3]. The second module, the explainer, feeds the top entities and their factor values from each component, along with decomposition quality metrics like reconstruction error and factor match score, into an LLM [4]. The LLM is guided by two types of prompts. Task-agnostic prompts instruct the model to describe the hidden patterns generally, while task-specific prompts direct it toward a particular goal [2]. The paper notes that task-specific explanations, such as generating movie recommendations from the MovieLens dataset, can be more actionable but may also introduce bias toward the given task [4]. To assess the reliability of the generated explanations, the researchers designed two evaluation protocols. In a forward inference task, the LLM must predict entities that should belong to a component but are absent from its top-ranked list, using a pool of the top-50 entities across all components to manage prompt length [3]. In a backward inference task, the model is asked to identify a false entity that has been deliberately injected into a component from a distinct cluster [4]. The broader context for this work is the exponential growth of complex datasets. Global data volume is projected to reach 163 zettabytes by 2025, and spending on big data and business analytics solutions was estimated at $215.7 billion in 2021 [6]. As artificial intelligence systems have advanced, particularly since the introduction of the transformer architecture in 2017, the ability to interpret the outputs of these models has become a critical field of study [5]. A demo of the AnTenA system is publicly available on GitHub [1].