Researchers built a skill bank so AI models stop reinventing bad math from scratch every time.
The framework, called OptiSkill, targets a specific weak spot: getting large language models to turn plain-English decision problems into correct mathematical programs, a task known as operations research modeling. Instead of solving each problem cold, OptiSkill stores solver-verified experience in a hierarchical, evolving SkillBank. That bank splits knowledge into Global Strategies, which are problem-level formulation skeletons, and Step Experiences, which are narrower rules for avoiding specific errors. New candidate skills only get added after they pass batch-level validation, so the bank should not accumulate bad habits over time. Across eight OR modeling benchmarks, the approach improved formulation accuracy across multiple LLM backbones and beat strong agentic baselines.
The real news is the shift from storing examples to storing reusable skills. Prior memory-based agents typically log past examples, thoughts, or insights as loose reference material. That works for recall but not for a task like OR modeling, where problems share underlying structure but rarely share surface wording, so a model needs transferable formulation know-how rather than a lookup table of similar cases.
It is a narrow fix for a narrow failure mode, not evidence that LLMs have gotten better at math reasoning generally. The interesting test will be whether SkillBank-style memory holds up outside curated benchmarks, on the messier optimization problems real businesses actually bring to these tools.