Source-linked AI summary
ReSum: Unlocking Long-Horizon Search Intelligence via Context Summarization
Xixi Wu, Kuan Li, Yida Zhao, Liwen Zhang, Litu Ou, Huifeng Yin, Zhongwang Zhang, Xinmiao Yu, Dingchu Zhang, Yong Jiang, Pengjun Xie, Fei Huang, Minhao Cheng, Shuai Wang, Hong Cheng, Jingren Zhou
TL;DR
Long-horizon web agents need extensive exploration, but accumulating interaction histories exhausts context windows and can make existing architectural solutions costly or incompatible. ReSum periodically compresses histories with an external summary tool, while ReSum-GRPO trains agents to reason over segmented summaries through advantage broadcasting. ReSum improves over ReAct in training-free settings, and ReSum-GRPO provides further gains while remaining data-efficient.
Problem
Complex web-search tasks require extended evidence gathering, but context limits truncate accumulated histories; prior architectural approaches can break compatibility and require end-to-end retraining.
Method
ReSum periodically condenses interaction histories into summaries, and ReSum-GRPO broadcasts trajectory-level advantages across segmented trajectories to train summary-based reasoning.
Results
ReSum achieves a 4.5% average improvement over ReAct, with ReSum-GRPO yielding a further 8.2% gain across three challenging benchmarks.
Takeaways & Limitations
With only 1K training samples, ReSum-GRPO enables a 30B agent to rival powerful open-source models trained on 10K+ samples.
Takeaways & Limitations
Future work identifies the need to replace rule-based external summarization with autonomous agent-driven mechanisms and add quality control for preserving long-range dependencies.
Abstract
from arXiv · showhide
Large Language Model (LLM)-based web agents excel at knowledge-intensive tasks but face a fundamental conflict between the need for extensive exploration and the constraints of limited context windows. Current solutions typically rely on architectural modifications, e.g., internal memory tokens, which break compatibility with pre-existing agents and necessitate costly end-to-end retraining. To overcome these limitations, we introduce ReSum, a lightweight, plug-and-play paradigm that enables unbounded exploration by periodically invoking an external tool to condense interaction histories into compact summaries. Although this paradigm functions without training, standard agents are not inherently aligned to reason over such compressed contexts. To bridge this gap, we propose ReSum-GRPO, which adapts Group Relative Policy Optimization (GRPO) via advantage broadcasting to propagate final rewards across segmented trajectories, enabling credit assignments over long-horizons. Extensive experiments show that ReSum achieves a 4.5% improvement over ReAct in training-free settings, with ReSum-GRPO yielding a further 8.2% gain. Notably, with only 1K training samples, a ReSum-enhanced 30B agent achieves competitive performance with leading open-source models, showing ReSum's effectiveness.
1. Introduction
ReSum addresses the conflict between long-horizon web exploration and limited context windows by periodically compressing interaction histories into summaries. The paper further introduces trained support for summary-based reasoning and reports gains over ReAct.
- Motivation: Complex web queries require extended cycles of targeted searching, browsing, and cross-verification to assemble evidence chains.
- Motivation: ReAct can exhaust limited context windows by appending every thought, action, and observation, while architectural memory solutions reduce compatibility and require costly retraining.
- ReSum: ReSum periodically invokes an external tool to condense interaction histories into compact summaries, enabling unbounded exploration without architectural modification.
- Supporting Tool: ReSumTool-30B is trained to extract key evidence, identify information gaps, and guide future search steps.
- ReSum-GRPO: ReSum-GRPO segments long trajectories and broadcasts final-outcome advantages across segments to adapt agents to summary-based reasoning.
- Results: 4.5% average improvement over ReAct was observed for ReSum, with a further 8.2% improvement after ReSum-GRPO training across three challenging benchmarks.
2. Preliminary
ReAct performs web exploration through iterative thought, action, and observation cycles conditioned on accumulated context. For difficult searches, this history can exceed the context window and force truncation before resolution.
- ReAct Workflow: ReAct iteratively generates a thought, emits a tool action, receives an observation, and repeats until producing a final answer.
- ReAct Workflow: A ReAct trajectory records the question followed by successive thoughts, actions, and observations across T iterations.
- ReAct Workflow: At step t, the policy samples the thought and action conditioned on all previous context.
- Context Limitation: Failed BrowseComp trajectories frequently exceed WebSailor-7B’s 32k context limit and require truncation, whereas solved cases generally fit within the window.
- Tools: The web-exploration implementation uses Search for Google results and Visit for URL-based browsing with goal-specific evidence extraction.
3. Methodology
ReSum periodically compresses interaction histories into goal-oriented summaries, allowing agents to continue long-horizon exploration within limited context windows. ReSum-GRPO adapts reinforcement learning to segmented trajectories by broadcasting a trajectory-level advantage across segments.
- ReSum Paradigm: ReSum periodically invokes an external summary tool to condense interaction histories, then resumes exploration from a compressed state.The compressed state combines the original query with a goal-oriented summary, while the working history is reset.
- ReSum Paradigm: Summarization can occur at systematic triggers such as token limits, while practical deployments still impose tool-call budgets.Trajectories exceeding those resource limits are terminated and marked as failures.
- ReSum Paradigm: ReSum transforms linear interaction histories into compact, restartable states while remaining functionally compatible with existing ReAct-based agents.This design supports long-horizon exploration that would otherwise overflow the context window.
- ReSumTool-30B: ReSumTool-30B is developed as a specialized summary model to extract key evidence, identify information gaps, and guide future search steps.Its development distills a teacher model’s capabilities into an efficient Qwen3-30B-A3B-Thinking-based model using Conversation–Summary pairs.
- ReSum-GRPO: ReSum-GRPO segments trajectories at summarization points and uses the final-answer reward to broadcast a normalized trajectory-level advantage across all segments.The approach modifies rollout collection and reward assignment while retaining standard processing for trajectories without summarization.
- ReSum-GRPO: Advantage broadcasting trains agents to reason from compressed states and gives earlier exploration steps credit when they contribute to final success.The mechanism uses a consistent learning signal per trajectory together with GRPO’s group-wise stabilization.
4. Experiments and Analysis
Experiments evaluate ReSum and ReSum-GRPO across challenging web-search benchmarks, agent scales, and training settings. ReSum improves training-free performance, while ReSum-GRPO further improves summary-conditioned reasoning and offers competitive results with limited training data.
- Experimental Setup: Experiments cover GAIA, BrowseComp, and BrowseComp-zh using WebSailor-3B, WebSailor-7B, and WebSailor-30B agents under a 32k-token context limit.Evaluation reports Pass@1 and Pass@3, with a maximum inference budget of 60 tool calls unless otherwise stated.
- Training-Free ReSum: ReSum achieves average absolute gains of 4.5% over ReAct across agents and benchmarks in training-free evaluation.The improvement is attributed to coherent exploration through intelligent context compression, whereas simple history truncation disrupts continuity.
- Training-Free ReSum: ReSumTool-30B reaches 13.7% Pass@1 on BrowseComp-zh with WebSailor-3B, exceeding Qwen3-235B at 11.1% and DeepSeek-R1-671B at 13.0%.These models are compared while serving as summary tools.
- Training-Free ReSum: WebSailor-30B with ReSumTool-30B achieves 16.0% Pass@1 on BrowseComp, surpassing Claude-4-Sonnet at 12.2% and Kimi-K2 at 14.1%.OpenAI-o3 and DeepSeek-v3.1 remain ahead in the reported comparison.
- ReSum-GRPO: ReSum-GRPO improves WebSailor-3B Pass@1 on BrowseComp-zh from 8.2% to 20.5% and standard GRPO does not transfer ReAct gains to ReSum.The results support paradigm-specific optimization for summary-conditioned reasoning.
- Efficiency and Adaptation: With only 1K training samples, ReSum-GRPO rivals open-source models trained on 10K+ samples, while MEM1-GRPO uses nearly 3× more tokens for a 1.2% improvement.ReSum-GRPO incurs approximately 1.5× training time because it preserves longer trajectories, but it can switch between ReAct and summary-based reasoning.
- Context Robustness: ReSum consistently outperforms ReAct across context configurations, including a 128k context, with larger gains under stricter context constraints.This indicates that the evaluated information-seeking tasks benefit from exploration beyond ordinary context limits.
5. Conclusion
The paper presents ReSum and ReSum-GRPO for long-horizon information-seeking tasks and reports effectiveness in training-free and training-required settings. It identifies autonomous summarization and quality control as future directions for preserving long-range dependencies.
- ReSum enables unbounded exploration through periodic summarization, while ReSum-GRPO provides training-based paradigm adaptation.
- Quality control is identified as a future direction for preserving long-range dependencies during summarization.
Impact Statement
The paper describes responsible-use commitments and situates ReSum among approaches for web-agent development and context management. It distinguishes ReSum’s context-compression objective from world-model-based forward planning.
- The work states that it uses no unauthorized datasets, test sets, or models and relies on publicly available or licensed data.
- Context Management for Agents: ReAct’s append-all-history strategy can cause unbounded context growth and rapid exhaustion on complex queries.
- Context Management for Agents: External components such as retrieval modules have been proposed to structure agent context more effectively.
- Distinction from World-model-augmented Agents: World models improve decision quality through dense forward-looking planning, whereas ReSum compresses prior context to support decision continuation.
- The paper provides an algorithmic description of ReSum and prompts that summarize conversation history before reasoning continues from the compressed context.
D.1. Implementation of Inference Paradigms
The inference paradigms are evaluated under a 32k-token context window and a common 60-call budget, with ReSum triggering external summarization near the context limit. RL configurations use shared hyperparameters while varying tool-call and token allocations.
- All WebSailor-series agents use a 32k-token context window, a maximum of 60 tool calls, temperature 0.6, and top p 0.95.
- ReAct: ReAct appends every thought, action, and observation, terminating as failure when the context window is reached without an answer.
- Recent History: Recent History retains only the recent 22k tokens when the context limit is reached, allowing the conversation to restart.
- ReSum: ReSum invokes ReSumTool-30B near the context limit to compress the conversation, using rule-based triggering for implementation efficiency.
- RL Training: All RL algorithms use batch size 64, group size 8, learning rate 2e−6, and 4 epochs under the 1K-sample setting.
- ReSum-GRPO: ReSum-GRPO allows 60 tool calls with a 4k-token prompt and 28k-token response allocation, invoking summarization when the token limit is reached.
E.1. Cases of LLMs in Context Summarization
The study compares GPT-OSS-120B with Qwen2.5-72B-Instruct for summarizing partially explored web-search histories. GPT-OSS-120B is reported to produce more structured, comprehensive, and goal-focused summaries, motivating its use in ReSumTool training.
- The evaluation uses a 2025 conference-deadline-and-CCF-rank question whose ground-truth answer is SIGIR 2025.
- The empirical study compares the context-summarization capabilities of GPT-OSS-120B and Qwen2.5-72B-Instruct.
- GPT-OSS-120B is reported to organize evidence by conference, explain candidate matches or failures, and maintain goal-oriented focus.
- ReSumTool-30B is trained from Qwen3-30B-A3B-Thinking using conversation-summary pairs collected with WebSailor-30B and GPT-OSS-120B.
E.3. Evaluation for ReSumTool-30B
ReSumTool-30B is evaluated through its effect on ReSum inference and through qualitative inspection of generated summaries. The results indicate useful performance gains and structured, goal-focused evidence organization.
- The illustrated summary concerns an IJCAI 2025 full-paper deadline of 23 January 2025, 23:59 AoE, and its CCF rank A.
- The evaluation combines quantitative ReSum inference performance in Table 1 with qualitative summary examples in Figure 6.
- ReSumTool integration provides comparable performance boosts to larger reasoning and instruction models while outperforming Qwen3-30B Base.
- ReSumTool-30B summaries exhibit reasonable structures, goal-focused organization, and comprehensive evidence gathering.
F. Supplementary Materials for Experiments
The supplementary materials extend the experimental analysis of ReSum-GRPO across training efficiency, inference costs, and representative cases.
- The supplementary experiments analyze ReSum-GRPO training efficiency, inference costs, and concrete trajectory cases.
F.1. Training Efficiency
ReSum-GRPO improves performance while adding measurable training and inference costs. Its resource use remains generally practical, and it compares favorably with MEM1 on token efficiency.
- Training Efficiency: Table 4 compares average time per gradient update step using batch size 64, G = 8, and 512 collected trajectories.
- Training Efficiency: ReSum-GRPO increases training time by approximately 33% to 69% compared with GRPO.The increase reflects longer trajectory collection and policy updates after segmentation at summarization points.
- Inference Costs: ReSum significantly boosts performance with only marginal resource-cost increases compared with training-free ReAct.Resource costs include average tokens and tool calls for successful trajectories.
- Inference Costs: ReSum paradigms typically maintain resource costs at approximately 2x the original costs while delivering substantial performance improvements.
- Inference Costs: ReSum-GRPO incurs additional resource costs because trained agents rely more on summaries for continued reasoning, but achieves higher performance.
- Inference Costs: ReSum-GRPO achieves comparable or higher success rates than MEM1-GRPO while consuming fewer tokens.MEM1-GRPO’s token overhead is attributed to iterative reasoning, memory consolidation, and tool invocation.
F.4. Case Study
The case studies illustrate both direct solving and reasoning resumed from summaries across challenging web-search questions. The examples show summaries preserving clues that support multi-step evidence gathering.
- Case Trajectories: ReSum-GRPO preserves the capability to solve simpler questions without summarization while supporting resumed reasoning on BrowseComp and BrowseComp-zh cases.
- Case Trajectories: The carnivorous-species case identifies Nepenthes khasiana by matching Meghalaya’s population density, conservation status, Indian genus representation, and local-name clues.
- Case Trajectories: A resumed BrowseComp trajectory uses a summary containing links among Shangganling, director Sha Meng, actress Ouyang Ruqiu, and her birthplace Xiaoxian County.
- Summary-Guided Reasoning: ReSumTool-30B is presented as extracting key clues and organizing evidence to guide subsequent search steps.
G. Justification for LLM-as-Judge Evaluation
The paper uses an LLM-based judge because ReSum produces open-ended responses, aligns with recent web-agent evaluations, and requires a reliable assessment protocol.
- The protocol uses Qwen2.5-72B-Instruct as the judge, supported by open-ended outputs, consistency with recent literature, and empirical reliability.
- LLM-based evaluation accommodates ReSum’s detailed, reasoned outputs better than strict Exact Match scripts designed for constrained formats.ReSum targets unbounded exploration and self-evolution, which naturally produces open-ended responses.
- Recent web-agent baselines use LLM-based evaluation for complex open-ended tasks, making the same protocol necessary for fair and direct comparison.The cited baselines include ASearcher, ARPO, and WebSailor.
- Four candidate judges showed high consensus on BrowseComp-zh, with score variance across models below 0.3%.The candidates were Qwen2.5-72B-Instruct, GPT-4o-Mini, Qwen3-235B, and Gemini-2.5-Flash.