Multimodal agents typically produce free-form reasoning before taking each action. According to a new paper posted on Hugging Face, this approach is inefficient for smaller models: limited capacity can lead to lengthy reasoning that offers little useful guidance for action generation while driving up inference costs.
To address this, the authors propose Selection-based Structured Reasoning (SSR), a framework that reframes reasoning as a selection task rather than open-ended generation. SSR encodes recurring high-level reasoning as pre-specified, reusable natural-language candidates. At each turn, the model selects among these candidates based on their likelihoods given the current context, without needing an auxiliary task head.
The use of pre-specified reasoning traces also enables parallel scoring: teacher-forced prefilling computes token likelihoods concurrently within and across candidates using a shared context KV cache.
The authors evaluate SSR on seven multimodal search benchmarks using 2B and 4B models. Across multiple reinforcement learning objectives and supervised fine-tuning, they report significant efficiency gains without sacrificing task performance. SSR achieves an average success rate competitive with leading search agents of the same scale, while reducing per-turn reasoning latency by over 90% and total per-question model inference latency by 28–54%.
A project page accompanies the paper at https://zfy0314.github.io/ssr-webpage/.