SSR reformulates the reasoning step of multimodal search agents as a selection among pre-specified natural-language candidates, rather than open-ended generation. At each turn, the model picks a candidate by its likelihood given the current context, using the language model itself as a classifier without an auxiliary head. This design allows parallel scoring of all candidates with a shared KV cache.
Evaluated on seven multimodal search benchmarks with 2B and 4B models, SSR achieves success rates competitive with leading trained search agents of the same scale. For example, a 4B SSR agent with GRPO training reaches 61.37% average success rate versus 61.25% for the baseline. At the same time, per-turn reasoning latency drops by over 90% and total per-question inference latency is reduced by 28–54%.
The key insight is that high-level reasoning in search tasks often repeats, so a small set of reusable candidates suffices. This makes SSR especially valuable for small, on-device models, where lengthy freeform reasoning provides little guidance while wasting computation. SSR maintains task performance while drastically cutting inference cost.