Proposal: a cuopt-multi-objective-exploration concept skill

[cafzal]

cuOpt optimizes one objective per solve. The formulation skill covers goal programming (fixed priorities), but nothing helps an agent when objectives genuinely trade off with no agreed weighting: there's no efficient-frontier exploration and no guidance for reading the tradeoff. That's common in many real-world business decisions, such as supply chain (e.g., cost vs. risk).

This concept skill turns repeated cuOpt LP/MILP/QP solves into a Pareto frontier (ε-constraint / weighted-sum sweep, then filter dominated) and adds the discipline to read it: quote the tradeoff, flag the knee, don't collapse to one "best." It adds no solver features and invents no API; it sits above the existing api-* and formulation skills.

I tested it with a with-vs-without A/B run with Claude models driving cuOpt as the solver, on a supplier-selection problem. With the skill, agents reliably present the tradeoff and defer instead of naming one answer, and they reach more of the frontier a weighted-sum sweep alone misses. It's a custom, LLM-judged A/B on synthetic instances – directional, not NVSkills-Eval.

• Notebook with eval results
• Staged skill folder (SKILL.md, evals, BENCHMARK, skill-card)

Would a skill like this be helpful, and where would you place it (common / API / new)? Happy to open a draft PR once you confirm.

[cafzal]

@anandhkb – you announced the cuOpt agent skills, so curious for your read. I've created similar optimization skills at RelationalAI and in my personal project Frontier, so would be glad to contribute this one.

[rgsl888prabhu]

@mlubin @chris-maes for viz

[rgsl888prabhu]

@cafzal This would be an awesome addition to the cuOpt skills, I think having this as it's own skill or may be a new type of skill and this is a reference underneath it would be a good spot. But I would like view from @mlubin and @chris-maes on the evaluation and technical details. But I think you can create a draft PR and we can follow-up on these over there as well.

May be to start with, I would suggest to create a new skill and iterate over other things in PR.

[cafzal]

Thanks @rgsl888prabhu @mlubin @chris-maes – I've opened a draft cuOpt skills PR for your review: #1355 , with a companion in the examples repo NVIDIA/cuopt-examples#151

Happy to discuss further!

[cafzal]

Separately: the cuopt-numerical-optimization-formulation skill still says quadratic constraints aren't supported (type table + "required questions" checklist), but the SOCP work (#1290) added convex quadratic constraints. Happy to file an issue or a small PR if helpful