Verified, then logged.
Sun 22 Nov 2026 · 00:16 KST / 15:16 UTC한국어 | ENSearch /

Breaking01 / AI

Apple's SSR cuts per-turn reasoning latency by more than 90%

The framework replaces freeform reasoning with a choice among preset candidates to reduce inference costs in small models, and Apple has not disclosed any product plans.

On a university lab bench, an oscilloscope connected to a palm-sized computer shows a single short pulse while an out-of-focus researcher, seen from behind, adjusts a probe. AI-generated image

Apple published a paper on 15 Oct 2026 reporting that Selection-based Structured Reasoning (SSR) cut per-turn reasoning latency in small models by over 90%, a company-published figure. SSR is a framework that replaces freeform reasoning with selection from a set of pre-specified candidates.

How it works. According to the paper, SSR writes out recurring high-level reasoning in advance as reusable natural-language candidates. At each turn, or action step, the model weighs each candidate's likelihood in the current context and selects one. Because the candidate text is fixed, token likelihoods were computed in parallel on a shared context cache.

The paper evaluated 2B and 4B models on seven multimodal search benchmarks, which cover several input formats. Total model inference latency per question fell by 28–54%, also a company-published figure.

The paper said SSR raised efficiency without sacrificing task performance across multiple reinforcement learning objectives and supervised fine-tuning. It also described SSR's average success rate as competitive with leading search agents of the same scale.

The research starts from the freeform reasoning that multimodal agents generate before each action. According to the paper, this reasoning can grow long in small models while offering little guidance for the action and carrying a high inference cost. The author list includes researchers affiliated with Carnegie Mellon University.

Apple did not disclose plans to use SSR in products on the paper's introduction page.