Researchers address the inefficiency of search-augmented language models by formulating an instance-level search-routing problem that decides whether external evidence is needed to improve task success. They derive supervision by comparing no-search and forced-search outcomes to construct an oracle, which serves as both an evaluation criterion and a learning signal for training policies via supervised fine-tuning and preference optimization.
- Search routing macro-F1 on oracle-eligible examples improved from 0.7082 to 0.8235 for Gemma E2B.
- The same metric increased from 0.7053 to 0.8365 for Qwen3.5-4B.
- Learned policies reduce model-specific routing failures, with Gemma primarily learning no-search restraint and Qwen reducing missed searches.
- Residual UNSOLVED cases reveal heterogeneous bottlenecks involving model capacity, retrieval budget, evidence use, and policy behavior.
This approach allows models to better determine when search is beneficial versus when correction, clarification, or abstention would be more appropriate.