Think Before You Retrieve: Learning Test-Time Adaptive Search with Small Language Models

The introduction of Orion marks a significant advancement in the field of information retrieval, particularly for small language models. Traditional methods have struggled with the complexities of user queries, often relying on large language models that are costly and less efficient. Orion addresses these challenges by integrating synthetic trajectory generation and reinforcement learning, allowing models with 350M to 1.2B parameters to refine their search strategies dynamically. Notably, the 1.2B parameter model achieved a 77.6% success rate on the SciFact dataset, surpassing the 72.6% success rate of prior retrievers, demonstrating that effective retrieval performance can emerge from learned strategies rather than just model size. This innovation not only enhances retrieval capabilities but also suggests a paradigm shift in how smaller models can be trained to perform complex tasks, potentially democratizing access to advanced AI technologies.