[Paper] ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization

In a significant advancement for AI-powered web search agents, researchers from Alibaba Group’s Tongyi Lab have developed a novel approach to overcome one of the most persistent limitations of Large Language Models (LLMs). Their paper, “ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization,” published on arXiv in September 2023, introduces a paradigm that enables web agents to maintain extended exploration capabilities without hitting context window constraints. The researchers’ innovative solution delivers substantial performance improvements across multiple benchmarks, potentially transforming how AI systems conduct complex information searches.

At its core, ReSum addresses a fundamental bottleneck in the widely used ReAct paradigm, where every interaction with the environment (thoughts, actions, and observations) is continuously appended to the conversation history. This approach quickly exhausts the context window of even the most advanced LLMs—typically limited to around 32,000 tokens—preventing agents from completing complex search tasks that require extensive exploration. Instead of appending every detail, ReSum periodically compresses the growing interaction history into compact reasoning states through strategic summarization, allowing agents to continue exploring indefinitely while maintaining awareness of previously discovered information.

Consider a complex query about identifying a literary work based on fragmentary biographical details of its subject: “A painter, whose father died of heart disease, had an elder sister and five children with his wife. Later, his marriage broke down and he had three more relationships. What is the name of the literary work based on this person?” Such questions demand multiple cycles of targeted querying, browsing, extraction, and verification to progressively reduce uncertainty and build a complete evidence chain—a process that traditional approaches often fail to complete due to context constraints.

The research team, led by Xixi Wu, Kuan Li, and Yida Zhao, implemented ReSum as a minimal modification to ReAct, ensuring simplicity and compatibility with existing agents. When approaching context limits, ReSum invokes a specialized summary tool that compresses the conversation into a structured format highlighting key discoveries and information gaps. This compressed state then serves as the new starting point, creating a seamless transition between exploration segments.

To power this summarization capability, the team developed ReSumTool-30B by fine-tuning Qwen3-30B-A3B-Thinking using carefully curated conversation-summary pairs. Unlike generic summarization tools, ReSumTool-30B specializes in extracting critical evidence from lengthy web search interactions, identifying information gaps, and highlighting next-step directions. This specialization allows the model to outperform even much larger models like Qwen3-235B and DeepSeek-R1-671B in summarization quality while maintaining deployment efficiency.

Furthermore, the researchers introduced ReSum-GRPO, a reinforcement learning framework that helps agents adapt to the ReSum paradigm. This algorithm extends Group Relative Policy Optimization (GRPO) to handle segmented trajectories, broadcasting trajectory-level advantages across all segments within the same rollout. This approach encourages agents to both reason effectively from compressed states and collect information that produces high-quality summaries. For those interested in learning more about reinforcement learning in AI, Reinforcement Learning: An Introduction provides an excellent foundation on these concepts.

The experimental results are impressive across multiple benchmarks. On the GAIA benchmark, WebSailor-30B achieved a 47.3% Pass@1 rate using ReSum with ReSumTool-30B, compared to 45.0% with standard ReAct. More dramatic improvements were seen on the challenging BrowseComp-en benchmark, where the same model configuration increased performance from 12.8% to 16.0%. When further enhanced with ReSum-GRPO training, the resulting WebResummer-30B model reached 18.3% on BrowseComp-en and an impressive 33.3% on BrowseComp-zh—surpassing several existing open-source web agents trained on much larger datasets.

These performance gains are particularly notable considering that ReSum-GRPO was implemented with just 1,000 training samples, compared to the 10,000+ samples typically used for training state-of-the-art web agents. The researchers attribute this efficiency to the paradigm’s ability to extend exploration capabilities without introducing excessive architectural complexity.

Beyond its immediate applications in web search, the ReSum approach represents a significant step toward developing more capable agentic systems. By enabling long-horizon reasoning without context constraints, the technique could benefit any domain requiring extended tool interaction and information gathering, from scientific research to complex planning tasks. For practical insights on building AI agent systems, AI Agents Unleashed offers valuable strategies for automating workflows and scaling business processes.

The research also points to promising future directions, including developing agents capable of intelligently self-initiating summary calls rather than relying on rule-based triggers, and further refining the summarization process to maximize information retention while minimizing token usage.

As LLM-based agents continue to evolve, approaches like ReSum that address fundamental architectural limitations without sacrificing simplicity or compatibility will play a crucial role in expanding their capabilities. For those interested in understanding the foundations and applications of large language models, Quick Start Guide to Large Language Models provides practical strategies and best practices. By enabling truly long-horizon exploration, these advances bring us closer to AI systems that can tackle increasingly complex, knowledge-intensive tasks with human-like persistence and thoroughness.