LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents (May 2026)
FreeContext-ReAct paradigm with five atomic operations (Skip, Compress, Rollback, Snippet, Delete) for adaptive context management; proves expressive completeness of Compress; LongSeeker achieves 61.5% on BrowseComp and 62.5% on BrowseComp-ZH, substantially outperforming Tongyi DeepResearch and AgentFol
About LongSeeker: Elastic Context Orchestration for Long-Horizon Search Agents (May 2026)
LongSeeker is a long-horizon search agent that introduces the Context-ReAct paradigm for adaptive context management. It provides five atomic operations (Skip, Compress, Rollback, Snippet, Delete) to dynamically reshape working context, preserving important evidence, summarizing resolved information, and discarding unhelpful branches. The Compress operator is proven to be expressively complete. LongSeeker is fine-tuned from Qwen3-30B-A3B on 10k synthesized trajectories and achieves state-of-the-art results: 61.5% on BrowseComp and 62.5% on BrowseComp-ZH, substantially outperforming existing agents like Tongyi DeepResearch and AgentFold.
Key Features
Pros & Cons
- State-of-the-art performance on BrowseComp benchmarks
- Provably expressive context management
- Reduces hallucination risk and generation cost via efficient operators
- Open-source and accessible via arXiv
- Only evaluated on synthetic trajectories and specific benchmarks
- Requires large base model (30B parameters)
- Not a ready-to-use product, requires implementation