How to Train Your LLM Web Agent: A Statistical Diagnosis

TL;DR

This paper presents a statistically grounded method for efficiently training LLM web agents, combining supervised fine-tuning and reinforcement learning to improve performance while reducing compute costs.

Contribution

It introduces a hyperparameter sampling and bootstrapping approach to optimize training strategies, achieving better performance with less compute compared to traditional methods.

Findings

Combining SFT with on-policy RL outperforms individual approaches.

The proposed method reduces compute by 45% while maintaining peak performance.

It effectively closes the gap with closed-source models.

Abstract

LLM-based web agents have recently made significant progress, but much of it has occurred in closed-source systems, widening the gap with open-source alternatives. Progress has been held back by two key challenges: first, a narrow focus on single-step tasks that overlooks the complexity of multi-step web interactions; and second, the high compute costs required to post-train LLM-based web agents. To address this, we present the first statistically grounded study on compute allocation for LLM web-agent post-training. Our approach uses a two-stage pipeline, training a Llama 3.1 8B student to imitate a Llama 3.3 70B teacher via supervised fine-tuning (SFT), followed by on-policy reinforcement learning. We find this process highly sensitive to hyperparameter choices, making exhaustive sweeps impractical. To spare others from expensive trial-and-error, we sample 1,370 configurations and use bootstrapping to estimate effective hyperparameters. Our results show that combining SFT with on-policy RL consistently outperforms either approach alone on both WorkArena and MiniWob++. Further, this strategy requires only 55% of the compute to match the peak performance of pure SFT on MiniWob++, effectively pushing the compute-performance Pareto frontier, and is the only strategy that can close the gap with closed-source models.