Researchers from Carnegie Mellon University and other institutions introduced Selection-based Structured Reasoning (SSR), a framework designed to reduce the reasoning cost of multimodal search agents built with small models. The paper was published on October 1 and submitted to Hugging Face on October 7 by Xiaoyu Zhu. It says small models can produce lengthy free-form reasoning before each action, even when that reasoning offers little guidance for choosing the next action.

SSR turns reasoning into a choice among prewritten, reusable natural-language candidates. At each turn, a model selects a candidate according to its likelihood in the current context, without an auxiliary task head. The researchers use teacher-forced prefilling to score candidates in parallel, calculating token likelihoods within and across candidates while sharing a context KV cache (stored attention information reused during generation).

The team tested SSR on seven multimodal search benchmarks with 2B and 4B models, using multiple reinforcement learning objectives and supervised fine-tuning. SSR maintained an average success rate competitive with leading search agents of the same scale. It reduced per-turn reasoning latency by over 90% and total model inference latency per question by 28-54%, while preserving task performance, according to the paper's abstract. The authors describe the gains as applying across the training approaches they evaluated.

Researchers from Carnegie Mellon University and other institutions introduced Selection-based Structured Reasoning (SSR), a framework designed to reduce the reasoning cost of multimodal search agents built with small models. The paper was published on October 1 and submitted to Hugging Face on October 7 by Xiaoyu Zhu. It says small models can produce lengthy free-form reasoning before each action, even when that reasoning offers little guidance for choosing the next action.

SSR turns reasoning into a choice among prewritten, reusable natural-language candidates. At each turn, a model selects a candidate according to its likelihood in the current context, without an auxiliary task head. The researchers use teacher-forced prefilling to score candidates in parallel, calculating token likelihoods within and across candidates while sharing a context KV cache (stored attention information reused during generation).

The team tested SSR on seven multimodal search benchmarks with 2B and 4B models, using multiple reinforcement learning objectives and supervised fine-tuning. SSR maintained an average success rate competitive with leading search agents of the same scale. It reduced per-turn reasoning latency by over 90% and total model inference latency per question by 28-54%, while preserving task performance, according to the paper's abstract. The authors describe the gains as applying across the training approaches they evaluated.