Decision models improve task success with small formalizer module
Learning to Formalize: Bridging Open-Ended Workloads and Fast Decision Models
Artificial Intelligence
Summary
Tasks often require quick decisions from a set list of options, but many real-world problems don’t offer clear choices upfront. The authors designed a small helper called a formalizer that creates specific possible options from an open-ended task description. This makes it easier for a decision model to quickly pick a good choice, improving success rates from about 12% to over 50% and reducing response time compared to large general models. Their approach bridges the gap between flexible language models and fast decision-making systems.
What this means in practice
- •For robotics engineers: Improve robot task planning by generating validated candidate actions from open-ended commands for faster and more reliable decision making.
- •For automation system developers: Enhance automated workflows by integrating a formalizer to produce actionable options from loosely specified tasks, reducing response time.
Authors
Chengxuan Pei, Ziyao Wang, Teng Mei, Zhe Li
Abstract
Decision models offer fast choices over sup- plied alternatives, while general-purpose large language models interpret open-ended tasks through flexible generation and reasoning at the cost of sequential decoding. Many work- loads expect a generative backend to determine what can be done and therefore provide no well-defined candidate list. We bridge this in- terface gap with a small candidate formalizer trained on offline teacher-generated lists. The formalizer constructs explicit, validated alter- natives from task context, and a frozen deci- sion model selects an entry. A motivating com- parison with official TypeSafe Jev-1.13.0 on 200 CALVIN-D test episodes reports 11% suc- cess with generic motions versus 46.5% with DeepSeek V4 Flash pose candidates. Sepa- rately, a six-arm privileged-state evaluation on 1,011 CALVIN-D validation annotations finds that a locally deployed decision model can achieve at most 52.6% task success rate with the fine-tuned formalizer, compared with 11.8% without it, and the end-to-end latency is kept to 2.09 s, whereas Qwen3.5-27B is 3.67 s with the exact experiment setting.