ReSum - Alibaba Tongyi's open-source WebAgent inference paradigm

What is ReSum?

ReSum, a novel WebAgent reasoning paradigm launched by Alibaba's Tongyi, belongs to the Tongyi DeepResearch family. It addresses the context length limitation problem faced by WebAgents in long-view-of-fact tasks by periodically summarizing the interaction history, compressing the ever-growing dialogue content into a compact reasoning state, enabling infinite exploration while maintaining awareness of previous discoveries. The core of ReSum is the ReSum-GRPO scheme, which integrates the GRPO algorithm, enabling agents to skillfully master summary-based conditional reasoning. Experiments show that ReSum performs exceptionally well across multiple tasks, achieving an average absolute performance improvement of 4.5% compared to the traditional ReAct method. This innovative mechanism provides broader possibilities for the application of WebAgents in complex tasks.

ReSum's main functions

• Breaking the context length limitBy using a periodic summarization mechanism, long dialogues are compressed into compact reasoning states, effectively solving the problem of context length limitations.
• Achieve infinite explorationReSum enables WebAgent to conduct unlimited exploration, dynamically update the inference state, and ensure that every decision is based on the latest information.
• Improve reasoning abilityReSum integrates the ReSum-GRPO scheme, which significantly improves the reasoning ability of intelligent agents through four steps: generation, retrieval, planning, and optimization.
• Experimental verificationExperiments show that ReSum achieves an average absolute improvement of 4.5% compared to the traditional ReAct method, and performs exceptionally well in long dialogues and complex tasks.

ReSum's technical principles

• Periodic context summarizationReSum compresses lengthy dialogues into compact reasoning states by periodically summarizing the interaction history, overcoming the limitations of context length while retaining key information to support subsequent reasoning.
• ReSum-GRPO algorithmReSum is based on the ReSum-GRPO scheme and integrates the GRPO (Group Relative Policy Optimization) algorithm. Through four steps of generation, retrieval, planning and optimization, it optimizes the decision-making process of the agent and improves its performance in complex tasks.
• Dynamic reasoning state updateReSum can dynamically update the inference state, ensuring that the agent infers based on the latest information at each step, thus achieving efficient multi-step inference.
• Reinforcement learning frameworkReSum is trained within a reinforcement learning framework. Through on-policy training and a customized GRPO algorithm, it ensures that the learning signal is always relevant to the model's current capabilities, thereby improving the stability and efficiency of training.

ReSum project address

• GitHub repository: https://github.com/Alibaba-NLP/DeepResearch/tree/main/WebAgent/WebResummer
• arXiv technical paper: https://arxiv.org/pdf/2509.13313

Application scenarios of ReSum

• academic researchIt efficiently processes complex academic literature and multi-step reasoning tasks, helping researchers quickly locate key information and improve research efficiency.
• Legal ResearchSystematically retrieve case law and cross-reference regulations to provide precise legal research support for legal professionals.
• Travel planningReSum can generate complex travel plans, such as multi-day road trip routes, including specific attractions and pet-friendly hotels, providing users with personalized travel suggestions.
• Medical consultationIt integrates patient medical records and the latest research to provide doctors with comprehensive medical information support and assist in the development of treatment plans.
• Financial AnalysisIt analyzes a large amount of financial data and market dynamics to provide investors with in-depth market analysis and investment advice.