FML-bench: A Controlled Study of AI Research Agent Strategies from the Perspective of Search Dynamics

2026-05-17

AI Summary

This work addresses the lack of a benchmark in existing AI research agents that decouples policy from execution infrastructure and enables process-level behavioral analysis. To this end, the authors propose FML-Bench, a comprehensive benchmark encompassing 18 foundational machine learning tasks across 10 domains, along with 12 process-level behavioral metrics to systematically evaluate diverse search strategies—including greedy hill climbing, tree search, and evolutionary optimization. The study innovatively decouples policy from execution and introduces an adaptive exploration mechanism based on improved opportunity density. Experimental results demonstrate that policy complexity does not directly correlate with performance; the proposed adaptive strategy switching significantly outperforms six baselines; and early convergence combined with directed exploration critically influences final performance, thereby confirming a strong association between exploratory behavior characteristics and overall efficacy.

Abstract

AI research agents accelerate ML research by automating hypothesis generation, experimentation, and empirical refinement. Existing agent strategies range from greedy hill-climbing to tree search and evolutionary optimization, yet which strategy choices drive performance remains unclear. Answering this question requires a benchmark that separates agent strategy (e.g., search topology) from execution infrastructure (e.g., code editor), so that performance differences are attributable to strategy rather than infrastructure, and that provides process-level metrics beyond final scores to analyze exploration behaviors. Existing benchmarks offer limited support. We propose FML-Bench, a benchmark of 18 fundamental ML research tasks across 10 domains that separates agent strategy from execution infrastructure and defines 12 process-level behavioral metrics. Evaluating six representative agents, we find that: (1) strategy complexity alone does not guarantee strong performance: a simple greedy hill-climber nearly matches the best-performing tree-search agent, both well above the remaining agents; (2) our analysis suggests this pattern relates to improvement opportunity structure: greedy search tends to be more effective when opportunities are dense, while tree-search and evolutionary strategies tend to be more effective when opportunities are sparse; an adaptive agent built on this insight switches to broader exploration upon detecting improvement stagnation and outperforms the other six agents, lending initial support to this observation; and (3) process-level analysis reveals that early convergence and directionally focused exploration are significantly associated with final performance, while solution diversity and compute cost are not. Our benchmark is available at: https://github.com/qrzou/FML-bench.

Problem

• AI research agents
• search dynamics
• strategy evaluation
• benchmarking
• process-level metrics

Innovation

• research agents
• search dynamics
• benchmark design
• adaptive exploration
• process-level metrics