Selection-Based Structured Reasoning: Toward Efficient Multimodal Search Agents

Researchers from Carnegie Mellon University and University of Illinois Urbana-Champaign introduced Selection-based Structured Reasoning (SSR), a method that improves efficiency of multimodal search agents by selecting from pre-specified reasoning candidates rather than generating freeform reasoning. SSR significantly reduces reasoning latency by over 90% and total inference time by up to 54%, while maintaining competitive task success rates on benchmarks using 2B and 4B parameter models.