Selection-Based Structured Reasoning: Toward Efficient Multimodal Search Agents
SSR turns agent reasoning into a selection problem, cutting per-turn reasoning latency by more than 90%.
The paper targets small multimodal search agents that waste inference on long free-form reasoning before actions. Its SSR framework uses reusable natural-language reasoning candidates and has the model score and select among them instead of generating a fresh trace. The authors report evaluations on seven multimodal search benchmarks with 2B and 4B models. They say SSR keeps success rates competitive with same-scale search agents while reducing total per-question model inference latency by 28-54%. HF Daily Papers' note
The paper targets small multimodal search agents that waste inference on long free-form reasoning before actions. Its SSR framework uses reusable natural-language reasoning candidates and has the model score and select among them instead of generating a fresh trace. The authors report evaluations on seven multimodal search benchmarks with 2B and 4B models. They say SSR keeps success rates competitive with same-scale search agents while reducing total per-question model inference latency by 28-54%. HF Daily Papers' note
score 4