ST-Bench: A Spatial-Temporal Benchmark for Multi-Agent System Generation on Scientific Research Tasks

2026-10-06

AI Summary

This study investigates whether multi-agent systems outperform single-agent approaches cost-effectively in complex scientific data analysis. We construct the first spatiotemporal benchmark for Earth sciences, leveraging LLM-driven multi-agent collaboration, automated workflow generation, and expert-validated datasets to systematically evaluate various generative methods on real-world analytical tasks regarding performance and computational overhead. Results demonstrate that most configurations exceed baseline scores by over threefold while significantly improving data coverage. Although high-performance configurations require approximately four times the execution time, cost-efficient variants capture the majority of performance gains at substantially lower computational expense. These findings provide a quantitative basis for navigating performance-cost trade-offs in deploying large language model-based agents for scientific data analysis.

Abstract

The rapid progress of LLM-based multi-agent systems (MAS) has shown that they largely outperform single agents on coding, math, and QA tasks, where executable tests provide a binary success signal. Whether this advantage transfers to real scientific data analysis remains untested. We introduce ST-Bench, a benchmark designed to answer two questions: whether MAS outperform single agents on complex scientific data analysis tasks, and if so, by how much and at what additional cost. ST-Bench contains 100 data science tasks adapted from published Earth science studies across hydrology, agriculture, and wetland methane research, expanded into 2,067 queries grounded in additional published studies and validated by domain experts. Using ST-Bench, we evaluate five recent MAS generation methods under two training protocols, against single-agent baselines on the same GPT-5 backbone. Nine of the ten MAS configurations exceed the cheapest single-agent baseline, with the strongest reaching nearly three times its composite score. This gain is primarily attributable to coverage: trained workflows produce realistic numerical metrics on a larger fraction of queries, while the quality of those metrics, conditional on producing realistic output, is comparable to that of the single-agent baseline. The strongest configuration requires approximately four times the single-agent inference time, whereas a more economical workflow captures the majority of the benefit at less than twice the cost. MAS specialization confers measurable benefit on scientific data analysis, but the benefit is conditional rather than universal.

Problem

• multi-agent systems
• scientific data analysis
• benchmark evaluation
• large language models
• spatial-temporal tasks

Innovation

• Multi-Agent Systems
• Benchmark
• Scientific Data Analysis
• Spatial-Temporal
• Large Language Models