Selection-Based Structured Reasoning (SSR): selection-based reasoning for efficient multimodal search agents
Authors introduce Selection-Based Structured Reasoning (SSR), a framework that reformulates agent reasoning as selection from pre-specified natural-language candidates rather than open-ended generation; SSR uses teacher-forced prefilling to enable parallel scoring and was evaluated with 2B and 4B models on seven multimodal search benchmarks. The paper reports SSR achieves competitive success rates while reducing per-turn reasoning latency by over 90% and total per-question inference latency by 28–54%.