SSR turns multimodal search-agent reasoning into candidate selection, cutting per-turn reasoning latency by over 90% while matching same-scale agents on 2B and 4B models
Related research and updates
Synopsis
The work introduces Selection-based Structured Reasoning (SSR), which reformulates the free-form reasoning that multimodal agents generate before each action as a selection over pre-specified, reusable natural-language reasoning candidates scored by their likelihood given the current context, without an auxiliary task head, using teacher-forced prefilling and a shared context KV cache for parallel scoring; on seven multimodal search benchmarks with 2B and 4B models, across multiple reinforcement learning objectives and supervised fine-tuning, SSR reaches an average success rate competitive with leading search agents of the same scale while reducing per-turn reasoning latency by over 90% and total per-question model inference latency by 28-54%.
Figure 1: SSR significantly reduces inference latency. Left: Parallel reasoning decoding replaces autoregressive reasoning generation. Right: Across four training objectives and two model sizes, SSR achieves comparable or higher success rates while reducing mean reasoning latency per turn by more than 90 % 90\% . Mean total model inference latency per question is reduced by 28 28 – 54 % 54\% (Table 2 ).
arXivInterpretation
SSR reformulates reasoning from open-ended generation into selection: recurring high-level reasoning is represented as pre-specified, reusable natural-language candidates, and at each turn the model selects among them based on their likelihoods given the current context. Relative to the common practice of freely generating reasoning before each action, the framework no longer has the model produce reasoning text token by token but restricts the reasoning space to a predefined candidate set. The abstract supports this design with a framework description and evaluation on seven multimodal search benchmarks, and notes that the selection process requires no auxiliary task head.
Pre-specified reasoning traces make parallel scoring possible: teacher-forced prefilling computes token likelihoods concurrently within and across candidates using a shared context KV cache. This mechanism converts what would be serial reasoning generation into parallelizable likelihood computation, which is the direct source of the efficiency gains. The abstract describes the mechanism and reports over 90% reduction in per-turn reasoning latency and 28-54% reduction in total per-question model inference latency.
On small 2B and 4B models, SSR delivers significant efficiency gains across multiple reinforcement learning objectives and supervised fine-tuning without sacrificing task performance. The result targets the problem that limited model capacity can yield lengthy reasoning that gives little useful guidance for action generation, indicating that efficiency improvements need not come at the cost of success rate. The abstract reports evaluation on seven multimodal search benchmarks, with an average success rate competitive with leading search agents of the same scale.
Perspective
The result is aimed at the multimodal search agent setting and applies to tasks whose reasoning steps contain recurring high-level patterns that can be pre-specified as reusable natural-language candidates; the evaluation covers 2B and 4B scale models and is validated across multiple reinforcement learning objectives and supervised fine-tuning. For practitioners seeking to compress reasoning latency at a similar model scale while preserving task success rate, SSR offers an implementation path that replaces generation with selection, with parallel scoring relying on teacher-forced prefilling and a shared context KV cache.
The abstract does not state the size, source, or construction of the reasoning candidate set, nor per-benchmark success rates and latency measurement conditions, so how stable the efficiency gains are across task distributions and candidate designs remains an open question. The abstract also does not report comparisons with larger-scale agents, leaving the relative advantage of SSR as model capacity grows to further observation.