ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization

1. Problem Statement

Large Language Model (LLM) based web agents excel at knowledge-intensive tasks but face a fundamental conflict: the need for extensive, long-horizon exploration versus the limited context windows of current LLMs.

• The Bottleneck: The standard ReAct (Reason + Act) paradigm appends every thought, action, and observation to the context history. For complex queries requiring multiple search cycles, this quickly exhausts the token budget (e.g., 32k tokens), forcing the agent to truncate history or fail before finding the answer.
• Limitations of Existing Solutions: Current approaches like MEM1 or MemAgent rely on architectural modifications (e.g., internal memory tokens) that break compatibility with pre-trained agents and require costly end-to-end retraining.

2. Methodology

The authors propose ReSum, a lightweight, plug-and-play paradigm, and ReSum-GRPO, a reinforcement learning algorithm to adapt agents to this new paradigm.

A. The ReSum Paradigm

ReSum enables unbounded exploration by periodically compressing the interaction history into a compact summary.

• Mechanism:
  1. Standard Interaction: The agent operates normally (Thought $\to$ Action $\to$ Observation).
  2. Trigger: When the context approaches the token limit (systematic trigger), a Summary Tool ($\pi_{sum}$) is invoked.
  3. Compression: The tool condenses the entire history ($H_t$) into a goal-oriented summary ($s$).
  4. Reset: The history is reset to a compressed state $H'_t = (q, s)$, where $q$ is the original query and $s$ is the summary. The agent resumes reasoning from this compressed state.
• Benefit: This allows the agent to maintain awareness of prior discoveries while freeing up context space for indefinite exploration without architectural changes.

B. ReSumTool-30B (Specialized Summary Model)

Generic LLMs struggle to distinguish key evidence from noisy, lengthy histories. The authors developed ReSumTool-30B to address this:

• Architecture: Based on Qwen3-30B-A3B-Thinking.
• Training: Distilled from a powerful teacher model (GPT-OSS-120B) using a dataset of 9,000+ high-quality $\langle$Conversation, Summary$\rangle$ pairs derived from the SailorFog-QA benchmark.
• Capabilities: Specifically trained to extract verifiable evidence, identify information gaps, and propose actionable next steps, outperforming much larger models (e.g., DeepSeek-R1-671B) in summarization quality for web search contexts.

C. ReSum-GRPO (Paradigm Adaptation via RL)

While ReSum works without training, standard agents are not inherently aligned to reason over compressed contexts. The authors introduce ReSum-GRPO (Group Relative Policy Optimization) to optimize agents for this paradigm.

• Trajectory Segmentation: Long trajectories are segmented at summarization points into multiple episodes ($H^{(1)}, H^{(2)}, \dots$).
• Advantage Broadcasting:
  • Instead of designing per-segment rewards, a unified trajectory-level reward is computed based on the final answer correctness ($R \in \{0, 1\}$).
  • This reward is normalized within a group of rollouts to calculate an advantage ($\hat{A}$).
  • Crucially, this advantage is broadcast to all segments within the trajectory.
• Goal: This mechanism solves the credit assignment problem in long-horizon tasks, encouraging agents to both reason effectively from compressed states and collect information that leads to high-quality summaries.

3. Key Contributions

1. ReSum Paradigm: A plug-and-play solution that decouples context management from the agent's architecture, enabling indefinite exploration without retraining or architectural changes.
2. ReSumTool-30B: A specialized, efficient summary model that outperforms significantly larger models in extracting key evidence and guiding search steps.
3. ReSum-GRPO: A novel RL algorithm that adapts agents to summary-based reasoning via trajectory segmentation and advantage broadcasting, requiring only 1K training samples for significant gains.

4. Experimental Results

Experiments were conducted on three challenging benchmarks: GAIA, BrowseComp, and BrowseComp-zh.

• Training-Free Performance:

  • ReSum (using ReSumTool-30B) achieved a 4.5% average improvement over standard ReAct.
  • It outperformed MEM1 in training-free settings, where MEM1 often degraded performance due to incompatibility.
  • A 30B agent enhanced with ReSum achieved 16.0% Pass@1 on BrowseComp, surpassing proprietary models like Claude-4 (12.2%) and Kimi-K2 (14.1%).

• Training-Required Performance (ReSum-GRPO):

  • With only 1K training samples, ReSum-GRPO yielded an additional 8.2% gain over the training-free ReSum baseline.
  • WebSailor-30B with ReSum-GRPO reached 33.3% Pass@1 on BrowseComp-zh, rivaling open-source models trained on 10K+ samples (e.g., ASearcher-32B at 15.6%).
  • Efficiency: ReSum-GRPO achieved comparable or better performance than MEM1-GRPO while consuming ~3x fewer tokens, offering a better performance-efficiency trade-off.

5. Significance and Impact

• Scalability: ReSum proves that even agents with large context windows (e.g., 128k) benefit from summarization, as complex information-seeking tasks often exceed current context limits.
• Cost-Effectiveness: It offers a path to high-performance agents without the massive computational cost of training new architectures or the data inefficiency of collecting expert trajectories for supervised fine-tuning.
• Generalizability: As a plug-and-play paradigm, it can be applied to existing off-the-shelf agents immediately, making it a practical solution for deploying robust long-horizon web agents in real-world scenarios.

In conclusion, ReSum resolves the context bottleneck in web agents through intelligent compression and RL-based adaptation, demonstrating that effective long-horizon search does not require massive model retraining or architectural overhauls.