What's the open problem here?
I have been working on these benchmarks, building an own-lineage pipeline (near-axis construction, ALM descent, canonical confirmation) rather than refining public designs. I reached 0.5818 on simple_to_build, confirmed against the canonical evaluator. I would value the community's and the Proxima team's view on what is actually left to solve.
Three questions:
- What do you consider actually unsolved here? Where is the frontier?
- What kinds of optimisation pipelines are you hoping the community develops, as opposed to leaderboard-tuning?
- What currently does not work well, in terms of efficiency, reliability, or generality to real device constraints?
Hi @haikomatt . Thanks for this, and for building your own pipeline end to end rather than refining published designs. That is the kind of contribution we hoped the benchmark would attract.
Where the frontier is: honestly, we don't know what metric values are ultimately achievable. We have new internal reference values that achieve ~178 HVI for the MHD-stable problem, but those reference values have been also beaten by the most recent submissions. What we value most are methodological contributions, because we keep iterating internally on the problem definitions themselves — adding new metrics as objectives or constraints. Methods that let us re-optimize quickly under a changed definition are worth more to us than a record on a fixed one.
On pipelines: we have no preference for what they look like, but we agree that leaderboard tuning is not the ultimate goal. What we want are reliable optimization methods that get us from a new problem definition to good designs as fast as possible. We are discussing internally releasing new problem definitions aimed at targeted scientific questions, and methods that can map Pareto fronts cheaply would be especially valuable there. If you are willing to write up or open-source your pipeline, we would be glad to see it.
What does not work well: nothing that is outright broken, but one thing we are genuinely uncertain about is how diverse near-optimal solutions can be — whether distinct regions of design space map to similar metrics. If that freedom exists, it could be exploited for costs the current problem definitions do not account for. We would consider that an open question worth taking on.
Happy to continue here, and feel free to reach out to the team directly if you want to go deeper on any of this.