Apple's SSR turns agent reasoning into multiple choice, cutting latency over 90%

量子位 · wechat · 2026-10-16

Apple and CMU propose SSR (Selection-based Structured Reasoning): instead of generating free-form thinking token by token, a search agent picks one of 6 predefined reasoning options (reverse image search, text query, zoom-in, answer from evidence, read image directly, common sense), then freely generates the concrete action (query, crop coordinates).

Mechanism

The LLM itself scores each option (softmax sampling) — no extra classifier head or SFT warm start;
All options share the context KV cache and are scored in parallel; a 4B model's 156-token thinking took 1822ms vs 32.66ms for SSR;
Plugs directly into GRPO-style RL training.

Results (Qwen3-VL 2B/4B, 7 multimodal search benchmarks)

4B SSR hits 61.37% average success, on par with same-size SOTA; 2B reaches 51.26%, comparable to zero-shot 8B;
Per-turn reasoning latency drops over 90%, p95 at 0.07s (baselines 0.8-1.9s);
Ablations: choosing only indices instead of full semantic text costs 3.9 points; options growing 1→6 lifts success from 37.24% to 61.37%.

The paper frames this as a reorganization of reasoning space: strategic reasoning compresses into a compact option set while queries, tool args and answers stay free-form.

Original post →

Apple's SSR turns agent reasoning into multiple choice, cutting latency over 90%

More from coding & agent