Apple proposes Selection-Based Structured Reasoning to cut multimodal agent inference costs

Apple’s Machine Learning Research team describes Selection-Based Structured Reasoning (SSR), a framework designed to address inefficiencies in small multimodal agents where freeform reasoning can be lengthy and costly. SSR reformulates reasoning as selection from pre-specified, reusable natural-language candidates rather than open-ended generation.

The approach enables parallel scoring by leveraging teacher-forced prefilling to compute token likelihoods concurrently within and across candidates using a shared context KV cache, avoiding the need for an auxiliary task head.

SSR was evaluated on seven multimodal search benchmarks using 2B and 4B models, with results showing significant efficiency gains. Across multiple reinforcement learning objectives and supervised fine-tuning setups, SSR maintained competitive success rates compared to leading search agents of the same scale.

The framework reduces per-turn reasoning latency by over 90% and total per-question model inference latency by 28–54%, according to the authors. The work includes contributions from researchers affiliated with Carnegie Mellon University and the University of Illinois Urbana-Champaign, with some authors noting prior or partial affiliation with Apple.