Apple's SSR recasts agent reasoning as selection, cutting per-turn latency over 90%
Apple's machine learning research team published a paper, "Selection-Based Structured Reasoning: Toward Efficient Multimodal Search Agents," proposing SSR, a framework that replaces freeform reasoning before each multimodal agent action with a selection over pre-specified, reusable natural-language reasoning candidates scored by likelihood given the current context, without an auxiliary task head. Because the reasoning traces are pre-specified, the method uses teacher-forced prefilling to score token likelihoods in parallel across candidates via a shared context KV cache. The authors report that on seven multimodal search benchmarks with 2B and 4B models, across multiple reinforcement learning objectives and supervised fine-tuning, SSR matches the success rate of leading search agents at the same scale while cutting per-turn reasoning latency by over 90% and total per-question model inference latency by 28–54%.
If the results hold up, turning reasoning into selection rather than generation could substantially cut inference cost for small-model agents, a plausible route to cheap on-device multimodal search agents.
AppleApple Machine Learning ResearchCarnegie Mellon University