---
library_name: transformers
pipeline_tag: text-generation
license: mit
language:
  - en
base_model:
  - Qwen/Qwen3-30B-A3B-Thinking-2507
tags:
  - agent
  - open-source
  - OpenSeeker
  - deep-research
---

<div align="center">

  <h1>LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents</h1>

</div>

<div align="center">

[![Code](https://img.shields.io/badge/Code-LongSeeker-24292F?style=for-the-badge&logo=github&logoColor=white)](https://github.com/PolarSeeker/LongSeeker)
[![Paper](https://img.shields.io/badge/Paper-arXiv%3A2605.05191-B31B1B?style=for-the-badge&logo=arxiv&logoColor=white)](https://arxiv.org/abs/2605.05191)
[![Model](https://img.shields.io/badge/Model-30B--SFT-FFD21E?style=for-the-badge&logo=huggingface&logoColor=white)](https://huggingface.co/PolarSeeker/LongSeeker-30B-SFT)

</div>


**Update — May 27:** The model has been updated. Please use the latest version for evaluation and deployment.


**LongSeeker** is a long-horizon search agent that introduces Context-ReAct, a novel paradigm for elastic context orchestration. Unlike standard ReAct agents that passively accumulate observations, LongSeeker dynamically reshapes its working context using five atomic meta-operations: Skip, Compress, Rollback, Snippet, and Delete. This allows the agent to preserve critical evidence, summarize resolved information, discard unhelpful branches, and control context size—achieving reliable and efficient long-horizon reasoning.

![image](https://cdn-uploads.huggingface.co/production/uploads/67b4079145dc598e0f110530/VFgj96qF70hDChu-ermSw.png)

## Highlights

- **Strong long-horizon search performance**: LongSeeker achieves **61.5** on BrowseComp, **62.5** on BrowseComp-ZH, **78.0** on xbench-2505, and **77.7** on GAIA-text, demonstrating competitive capability across both web search and general agent benchmarks.
- **Elastic context orchestration for search agents**: We introduce **Context-ReAct**, a new agentic paradigm that jointly generates reasoning, context meta-operations, and tool calls, enabling agents to dynamically decide **when, where, and how** to reshape their working context during long-horizon search.
- **Comprehensive and fine-grained context control**: Context-ReAct defines five atomic operations—**Skip, Compress, Rollback, Snippet, and Delete**—forming an expressively complete yet efficient operation set for multi-resolution context management.
- **Efficient context management at extended horizons**: LongSeeker maintains a stable working context of around **15k tokens** even across long trajectories, using only a small fraction of its **256k** context window while avoiding the rapid context growth of standard ReAct agents.

## Performance


![image](https://cdn-uploads.huggingface.co/production/uploads/67b4079145dc598e0f110530/MvKon_63ikwQ6BbOjU4L2.png)

For more details, please refer to our [GitHub repository](https://github.com/PolarSeeker/LongSeeker).
Paper: [arXiv:2603.15594](https://arxiv.org/abs/2605.05191)