Back to AI Research

AI Research

ArchOR retrieves optimization skills by mathematical structure

Key Takeaways

  • ArchOR changes how an optimization agent selects existing workflow skills, using structural signatures and a fallback to the original selector.
  • ArchOR studies how an AI agent can reuse an optimization workflow when the wording of a task changes.
  • The [paper](https://arxiv.org/abs/2610.41909) builds on OptSkills and concentrates on selecting suitable skills from an existing library, rather than introducing a new solver or rebuilding the library.
  • CARL refines cached skill representations using archetype prototypes and a low-rank residual.
  • Signature-Guided Retrieval attaches structural descriptions to skills and uses weighted overlap to prioritize candidates before a bounded language-model reranking step.

ArchOR studies how an AI agent can reuse an optimization workflow when the wording of a task changes. The paper builds on OptSkills and concentrates on selecting suitable skills from an existing library, rather than introducing a new solver or rebuilding the library.

Retrieval based on the problem

The system adds two retrieval modules. CARL refines cached skill representations using archetype prototypes and a low-rank residual. Signature-Guided Retrieval attaches structural descriptions to skills and uses weighted overlap to prioritize candidates before a bounded language-model reranking step.
A structural description can capture the class of problem, the variable pattern and the family of constraints. These properties help distinguish requests that share words but require different mathematical formulations. They can also connect differently phrased requests that use the same underlying optimization structure.

A fallback remains part of the design

The arbitration rule preserves the original selector when the structural route lacks convincing support. That is a useful engineering detail: a new retrieval signal can improve some decisions while remaining unreliable on tasks outside its coverage.
With the same DeepSeek-V4.1-Flash backbone and released skill library, the paper reports 65.00% micro-average Pass@1 across five benchmarks. Its reported improvement over OptSkills is 1.64 percentage points on the micro average and 2.68 points on the macro average. These are results for the study's setup, rather than evidence that every business optimization problem can now be solved correctly.

How to apply the finding

For a modeling assistant, retain both the natural-language request and the formal structure used to select a skill. Test paraphrases that preserve the mathematics, alongside similar-sounding problems that change a constraint. Evaluate whether the solver's accepted output represents the intended business requirement. Solver checks help validate an executable formulation, but the formulation itself must still match what the user meant.

Comments