Apple Machine Learning Research published a paper on 15 October introducing Selection-Based Structured Reasoning (SSR), a framework designed to cut reasoning latency in multimodal search agents. Instead of generating open-ended reasoning before every action, SSR treats reasoning as a selection problem, drawing from a set of pre-specified, reusable natural-language candidates that can be scored in parallel.
The authors, a team from Apple and the University of Illinois Urbana-Champaign led by Feiyu Gavin Zhu, Xiaoyu Zhu and colleagues, evaluated SSR on seven multimodal search benchmarks using models with 2 billion and 4 billion parameters. The results show the framework matches the success rate of leading search agents at the same scale while cutting per-turn reasoning latency by more than 90 per cent. Total per-query model inference latency fell by 28–54 per cent. (The researchers’ paper does not name the specific benchmarks or the agents it was tested against.)
The work builds on Apple’s earlier research into efficient reasoning, including Interleaved Reasoning for LLMs via Reinforcement Learning and Learning Structured Reasoning via Tractable Trajectory Control.
Why it matters
Small models that produce long, freeform reasoning waste time and compute for little gain. SSR’s shift to a selection-based design lets a modest 2B or 4B-parameter model act decisively. Apple researchers have been pursuing similar efficiency gains in other agent contexts. In our view, SSR is a practical step towards running capable multimodal agents locally, where latency and compute are strict constraints.