FAMOSE: A ReAct Approach to Automated Feature Discovery

There is no information available in the provided context about a framework called FAMOSE or its association with the ReAct approach for automated feature discovery. The context does discuss automated feature engineering and the ReAct framework separately, but no direct link between them is established.

Automated feature engineering involves transforming raw data into meaningful features using systematic, algorithmic methods, reducing the need for manual intervention and domain expertise. Techniques such as deep feature synthesis (DFS), dimensionality reduction, and genetic algorithms are used to generate, transform, and select features automatically. Tools like FeatureTools implement DFS to create complex features from relational datasets by applying aggregation operations like mean, count, and max across related tables.

On the other hand, the ReAct framework, referenced in the context, refers to a reasoning-and-action paradigm used in large language models (LLMs) to improve decision-making through an iterative loop of reasoning, taking action (e.g., calling an API), and observing outcomes. It has been applied to enhance agent systems by reducing hallucination and enabling autonomous data annotation.

While both automated feature engineering and ReAct aim to reduce human intervention in AI workflows, there is no evidence in the provided context that FAMOSE exists or combines these two concepts. Therefore, claims about FAMOSE as a ReAct-based framework for automated feature discovery cannot be substantiated based on the current information

FAMOSE introduces a novel framework for automated feature engineering (AFE) in tabular machine learning by leveraging the ReAct (Reasoning + Acting) paradigm. Unlike traditional AutoML methods that rely on exhaustive search over predefined transformation libraries, FAMOSE utilizes Large Language Models (LLMs) to autonomously reason about the data schema and target variable. The system operates through an iterative loop where it generates hypotheses about useful features, executes code to create them, evaluates their impact on model performance, and refines its strategy accordingly. This approach allows the framework to handle the entire feature lifecycle—from generation and transformation to final selection—without requiring manual intervention or domain-specific heuristics.

A key contribution of this work is the integration of semantic understanding into the feature discovery process. By employing an LLM-based agent, FAMOSE can interpret the meaning of column names and data distributions, enabling it to construct features that are logically relevant rather than just mathematically derived. For instance, it can infer relationships between date formats or categorical variables that traditional statistical methods might miss. The authors demonstrate that this reasoning capability allows FAMOSE to outperform existing state-of-the-art AutoFE tools across various benchmarks, achieving higher predictive accuracy with a more efficient search process that minimizes computational redundancy.

The significance of FAMOSE lies in its potential to democratize high-performance machine learning by reducing the reliance on costly domain expertise. Feature engineering is often cited as the most labor-intensive part of the ML pipeline, acting as a bottleneck for data scientists. By automating this cognitive task through reasoning agents, FAMOSE not only accelerates model development but also uncovers non-obvious feature interactions that human experts might overlook. This represents a step forward in semantic AutoML, moving the field from brute-force optimization toward intelligent, context-aware system design.

FAMOSE: A ReAct Approach to Automated Feature Discovery

FAMOSE introduces a novel ReAct-based framework designed to automate the feature engineering pipeline for tabular machine learning (ML) data. The approach leverages a Reinforcement Learning + Acting (ReAct) paradigm to dynamically generate, refine, and select optimal features, eliminating the need for manual feature engineering—a process traditionally reliant on domain expertise. The system iteratively explores potential transformations (e.g., mathematical operations, aggregations, or interactions) and evaluates their utility using a learned reward signal derived from downstream ML performance. By integrating feature generation with model training, FAMOSE adapts to the data distribution, producing features tailored to the specific predictive task.

The paper’s key contributions include: 1. Automated Feature Synthesis: FAMOSE replaces heuristic-based or template-driven feature engineering with a data-driven, optimization-aware process, improving efficiency and scalability. 2. End-to-End Learning: The framework jointly optimizes feature generation and model training, unlike traditional pipelines where features are engineered in isolation. 3. Generalizability: Demonstrated effectiveness across diverse tabular datasets, with improvements in model performance and reduced manual effort.

This work is significant for ML practitioners and researchers, as it addresses a persistent bottleneck in tabular data workflows—feature engineering—by providing a scalable, automated solution. By reducing dependency on domain knowledge, FAMOSE lowers the barrier to entry for applying ML to new domains while enhancing reproducibility. The ReAct formulation also opens avenues for further exploration in automated ML (AutoML), particularly in dynamic or high-dimensional settings.

Source: [arXiv:2602.17641](https://arxiv.org/abs/2602.17641)