4B HybridResearcher beats larger Qwen model on cross-source research
Original title: HybridResearcher: Bridging the Structured-Unstructured Boundary for Deep Research Agents
It gains 30.0 pass@8 points over Qwen3.5-4B and 8.3 over Qwen3.5-397B-A17B on HybridDeepResearch.
HybridResearcher is a 4B agent trained to join structured database evidence with unstructured web evidence. Training data starts from validated single-source tasks and adds a second source around a bridge entity, kept only when neither side alone suffices. Training uses supervised fine-tuning followed by reinforcement learning. Authors also report gains versus Qwen3.5-4B on held-out BrowseComp and BrowseComp-Plus.
Unconfirmed: Benchmark margins are the authors' reported results, not an independent evaluation.
- Training starts from validated text-to-SQL or web-search tasks.
- A sample is kept only if neither source alone can answer.
- Held-out BrowseComp and BrowseComp-Plus scores also rose versus the base model.
From the original article
1 min read at huggingface.co
While LLM agents are increasingly capable of gathering and synthesizing evidence from a single homogeneous source, such as structured tables or unstructured data, real-world analytical questions frequently span heterogeneous sources. For example, answering one enterprise question may require both a result computed over an internal database and evidence found on the web.
By Soyoung Yoon, Peiran Xu, Han Wang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.