TL;DR
HybridResearcher-4B is a compact agent designed to solve complex analytical questions that require "bridging" the gap between structured databases (SQL) and unstructured web data. By using a novel data synthesis pipeline that forces cross-source interdependence, this 4B-parameter model outperforms frontier models 100x its size on hybrid research benchmarks while actually improving its performance on single-source tasks.
What Problem Does This Paper Address?
While Large Language Model (LLM) agents have become proficient at using specific tools—like writing SQL to query a database or searching the web—they struggle with tasks that require cross-source reasoning.
In the real world, an enterprise question might require you to first find a specific transaction ID in a database and then search the web to see if that transaction's vendor has recently filed for bankruptcy. Current agents often fail to carry evidence from one source to the next. The authors identify two main obstacles:
- Composition Gap: Training on single-source data (SQL-only or Search-only) does not automatically teach a model how to bridge them.
- Shortcut Learning: Models often bypass tools entirely, answering from their internal "parametric memory" instead of performing the necessary cross-source hop.
Methods
The authors introduce a training framework that explicitly teaches the "bridge" between sources.
1. Synthesizing Cross-Source Trajectories
Instead of building hybrid tasks from scratch, they start with validated "seeds" (existing text-to-SQL or web search tasks) and extend them across the boundary:
- SQL2S (SQL-to-Search): The agent executes a SQL query to find a "bridge entity" (e.g., a specific company name), which it then uses to perform a multi-step web search.
- S2SQL (Search-to-SQL): The agent resolves a web search to find a parameter (e.g., a specific year or ID) required to complete a complex SQL query.
To prevent "shortcuts," the authors use coined facts—inventing novel business rules or entities using pronounceable pseudo-words. This forces the model to actually use the tools because the information doesn't exist in its internal memory.
2. Training Recipe
- Supervised Fine-Tuning (SFT): The model is trained on high-quality trajectories generated by "teacher" models (like Qwen-397B).
- Reinforcement Learning (RL): To refine the policy, they use a technique called CISPO. They implement hard-gating (terminating rollouts that use tools inefficiently or redundantly) and judge majority-voting (using five LLM judges to provide a more stable reward signal).
Datasets
- Training: 21K synthesized hybrid tasks, filtered down to 0.7K high-signal pairs for RL. They also use existing datasets like HyDRA and various public text-to-SQL benchmarks (Spider, BIRD).
- Evaluation: The primary benchmark is HybridDeepResearch, which specifically tests the ability to combine SQL and search.
- Single-Source Benchmarks: To check for "catastrophic forgetting," they evaluate on Spider 2.0-lite (SQL) and BrowseComp/BrowseComp-Plus (Search).
Key Insights and Ideas
- The "Bridge" is the Skill: Simply mixing SQL and search data in training isn't enough. The model must be explicitly taught the interdependence—how the output of one tool conditions the input of the next.
- Anti-Shortcut Guarantees: By using "construction gates" (deterministic checks ensuring neither source alone can solve the problem), the authors ensure the model learns rigorous cross-referencing rather than lucky guessing.
- Efficiency via RL: RL doesn't just improve accuracy; it transforms the agent's behavior. Trained agents move from "blind searching" to "active reading," visiting more pages per search and using more specific, deliberate queries (e.g., using quotes and advanced search operators).
Key Results
- Superior Performance: HybridResearcher-4B improved over its base model (Qwen3.5-4B) by 30.0 points (pass@8) on the HybridDeepResearch benchmark.
- Size Doesn't Equal Skill: It surpassed the much larger Qwen3.5-397B by 8.3 points, proving that targeted training on reasoning "bridges" is more effective than sheer scale.
- Synergistic Effects: Remarkably, training on hybrid tasks improved single-source performance. SQL accuracy on Spider 2.0-lite rose from 59.26% to 65.19%, and web search performance also saw significant gains.
- Behavioral Shift: After RL, exact query repeats dropped by 63%, and the use of advanced search operators increased by 126%, indicating a much more "deliberate" research policy.
Limitations and Future Work
- Retrieval Bottlenecks: The agent is still limited by the underlying search engine. If the required bridge entity doesn't appear in the top search results, the reasoning chain breaks.
- Early Stopping: Many failures occur because the agent hits a time or turn budget and resorts to guessing.
- Future Directions: The authors suggest that the "bridge" blueprint could be extended to other heterogeneous sources, such as combining internal document repositories with code execution or domain-specific APIs.
Summary updated