LongAgent: History-Guided Agentic Search for Longit... | AI Research

LongAgent: History-Guided Agentic Search for Longitudinal Outcome Prediction
Medical research often relies on longitudinal data—measurements collected over time—to predict future health outcomes. However, these datasets are complex, containing many variables with different sampling rates and sources. Traditionally, researchers must manually select which variables to use, define the time windows for analysis, and choose how to aggregate the data, which is a time-consuming and often inefficient process. This paper introduces LongAgent, an AI-driven system designed to automate this search process by autonomously identifying the most predictive combinations of variables, time windows, and mathematical aggregation functions.

How LongAgent Works

LongAgent functions as an intelligent agent that treats feature discovery as a "budgeted search problem." Instead of relying on manual trial and error, the agent iteratively explores a space of potential candidates. Each candidate is defined by a specific set of variables, a time window relative to the outcome, and an aggregation method (such as mean or median). The ai search story also surfaces in Stanford AI discovery identifies natural weight..., adding another angle.
The agent uses a two-stage decision process to navigate this search space. It is guided by a "history memory" that integrates four types of evidence: the current search status, the history of previously evaluated candidates and their performance, a summary of which areas of the search space have already been covered, and evidence from past transitions (how changing a variable or window affected the prediction error). By analyzing this information, the agent decides whether to add, drop, or replace variables, or to update the time window and aggregation function.

Performance and Results

The researchers evaluated LongAgent using 360 synthetic datasets and a real-world clinical dataset from the IDEA-FAST study, which tracks fatigue scores in over 1,000 participants. The ai search story also surfaces in OpenAI Says AI Found Possible Navier–Stokes..., adding another angle.
On the synthetic datasets, LongAgent consistently outperformed three non-agent baseline methods, achieving the lowest mean prediction error (RMSE). An ablation study—a test to see how different parts of the system contribute to its success—revealed that providing the agent with "transition evidence" (information about how past changes impacted performance) significantly improved its ability to find better predictive features. On the real-world clinical dataset, LongAgent performed comparably to or slightly better than the strongest baseline method, demonstrating its practical utility in real medical research scenarios.

Limitations and Future Directions

While LongAgent shows promise in automating the discovery of predictive features, the current version has specific limitations. At present, the system applies a single, shared time window and aggregation function to all selected variables. The authors note that future research will focus on developing variable-specific temporal representation learning, which would allow the agent to tailor its analysis to the unique characteristics of different medical variables. Additionally, the team plans to expand the method to handle more diverse, heterogeneous, and incomplete longitudinal datasets. The ai search story also surfaces in New AI Architecture Mimics the Human..., adding another angle. as detailed in the full paper on Arxiv