LATTEArena: An Evaluation Framework for LLM-powered Tabular Feature Engineering (Extended Version)

Researchers introduce LATTEArena, a standardized evaluation framework for comparing LLM-powered tabular feature engineering methods. The framework decomposes 15 representative techniques into reusable components and reveals that Tree-of-Thought combined with Monte Carlo Tree Search offers optimal cost-effectiveness, while RPN and Code formats excel at different task types.

LATTEArena addresses a critical gap in AI research infrastructure by providing the first standardized platform for evaluating LLM-powered feature engineering approaches. The tabular data analysis domain has become increasingly complex as researchers integrate multiple advanced techniques—Tree-of-Thought, few-shot learning, Monte Carlo Tree Search, and natural language generation—into unified systems. Without comparative benchmarks, the field struggles to isolate which components actually drive performance gains versus adding unnecessary complexity and cost.

The framework's six-dimensional taxonomy and modular architecture enable controlled experimentation that wasn't previously possible. By decomposing 15 methods into reusable components and running over 4,000 execution logs, the researchers create a resource that eliminates the methodological opacity plaguing LLM-powered feature engineering research. This approach mirrors how benchmarking frameworks have accelerated progress in other AI domains.

For the broader AI ecosystem, LATTEArena demonstrates that standardization and cost-awareness are becoming central concerns as LLM applications mature. The finding that Tree-of-Thought with Monte Carlo Tree Search achieves optimal cost-effectiveness while RPN and Code formats dominate different task types provides actionable insights for practitioners. Organizations building production systems can now reference empirical evidence rather than heuristics when selecting feature engineering approaches.

The public release of the framework and execution logs creates a foundation for continuous improvement. Future researchers can systematically test novel techniques against established baselines, accelerating innovation cycles. This infrastructure-first approach suggests the field recognizes that progress increasingly depends on shared evaluation standards rather than isolated breakthroughs.

• →LATTEArena provides the first standardized competitive evaluation framework for LLM-powered tabular feature engineering methods.
• →Tree-of-Thought combined with Monte Carlo Tree Search achieves the best cost-effectiveness ratio across tested methods.
• →Component-level ablation studies quantify the isolated impact of individual techniques, revealing which contributions matter most.
• →RPN and Code output formats show task-specific dominance for classification and regression respectively.
• →Public release of 4,000+ execution logs enables researchers to benchmark new techniques against established baselines systematically.