How to Train Your Deep Research Agent? Prompt, Reward, and Policy Optimization in Search-R1

TL;DR

This paper systematically studies reinforcement learning strategies in Deep Research agents, revealing key insights on prompt templates, reward functions, and optimization methods, and introduces an improved baseline called Search-R1++.

Contribution

It provides a comprehensive analysis of RL components in Deep Research agents and proposes a new baseline that enhances performance based on these insights.

Findings

Fast Thinking template improves stability and performance.

F1 reward underperforms EM due to answer avoidance, mitigated by penalties.

REINFORCE outperforms PPO in efficiency and stability.

Abstract

Deep Research agents tackle knowledge-intensive tasks through multi-round retrieval and decision-oriented generation. While reinforcement learning (RL) has been shown to improve performance in this paradigm, its contributions remain underexplored. To fully understand the role of RL, we conduct a systematic study along three decoupled dimensions: prompt template, reward function, and policy optimization. Our study reveals that: 1) the Fast Thinking template yields greater stability and better performance than the Slow Thinking template used in prior work; 2) the F1-based reward underperforms the EM due to training collapse driven by answer avoidance; this can be mitigated by incorporating action-level penalties, ultimately surpassing EM; 3) REINFORCE outperforms PPO while requiring fewer search actions, whereas GRPO shows the poorest stability among policy optimization methods. Building on these insights, we then introduce Search-R1++, a strong baseline that improves the performance of Search-R1 from 0.403 to 0.442 (Qwen2.5-7B) and 0.289 to 0.331 (Qwen2.5-3B). We hope that our findings can pave the way for more principled and reliable RL training strategies in Deep Research systems.