Decision models can quickly select among a provided set of alternatives, whereas general‑purpose large language models handle open‑ended tasks through flexible generation and reasoning at the cost of sequential decoding latency. Many workloads lack a well‑defined candidate list and rely on a generative backend to determine feasible actions. To close this gap we train a small candidate formalizer on offline teacher‑generated lists. The formalizer builds explicit, validated alternatives from the task context, and a frozen decision model then picks one entry. In a comparison with the official TypeSafe Jev‑1.13.0 on 200 CALVIN‑D test episodes, generic motions achieve 11% success, while pose candidates generated by DeepSeek V4 Flash raise success to 46.5%. A further six‑arm privileged‑state evaluation on 1,011 CALVIN‑D validation annotations shows that a locally deployed decision model can reach up to 52.6% task success when paired with the fine‑tuned formalizer, versus only 11.8% without it. End‑to‑end latency stays at 2.09 s, compared with 3.67 s for Qwen3.5‑27B under the same experimental setting.
Review