HybridResearcher: Bridging the Structured-Unstructured Boundary for Deep Research Agents

HybridResearcher. Bridging the Structured-Unstructured Boundary for Deep Research Agents. By Soyoung Yoon and ten others, from Seoul National University, the University of California, Los Angeles, and three other institutions. Published on the seventeenth of October. No code repository is linked to it.

Agents built on large language models are increasingly good at gathering and synthesizing evidence from one homogeneous source. That source might be structured tables, or it might be unstructured data. Real analytical questions often refuse to stay in one place, though. A single enterprise question might need a result computed over an internal database and, next to it, evidence that can only be found on the web. Neither half answers the question without the other.

Recent evaluations on benchmarks like HybridDeepResearch show that even frontier agents struggle when evidence has to cross that boundary. So the difficulty has been measured. What has been missing is a method that explicitly teaches cross-source reasoning, rather than only exposing its absence.

HybridResearcher is a four-billion-parameter agent trained to bridge evidence across the structured and unstructured boundary. Its central ingredient is training data in which crossing the boundary cannot be avoided. Data like that is scarce, so the authors build it by synthesis. They start from tasks that already work on one side of the boundary, and they manufacture the other side around them.

Each synthesis begins from a validated single-source task, such as an existing text-to-SQL query or an existing web search query. From the seed task's own output, a bridge entity is drawn, and the missing source is fabricated around that entity. The result is a trajectory that has to pass through both kinds of evidence to reach its answer.

Not every synthesized sample survives. A sample is retained only when strict construction gates confirm that neither side alone is enough to answer it. That filter is what guarantees the data demands genuine interdependence between the sources, rather than a task one side could quietly solve on its own. The model is then trained on this curated set in two stages: supervised fine-tuning first, followed by reinforcement learning.

On the existing HybridDeepResearch benchmark, HybridResearcher improves over its own base model, Qwen3.5-4B, by thirty points pass at eight. It also surpasses the far larger Qwen3.5-397B-A17B, by eight point three points. On two established single-source benchmarks that were held out, BrowseComp and BrowseComp-Plus, it still improves over its base model.

That last result carries as much weight as the headline. The authors read it as evidence that complex cross-source reasoning can be taught directly to a small model without triggering catastrophic forgetting, so the new skill is added on top of the old ones rather than in place of them. A small model, given the right data, overtakes a far larger one on this task.