Researchers have built an AI optimization framework that changes its own search strategy mid-run instead of locking in one approach from the start.
The system, called OptiCom, comes from a new arXiv paper on LLM-driven optimization. It organizes every optimizer decision into six shared components: the artifact being improved, the query driving it, the operators available, how results get evaluated, what gets remembered, and the overall strategy. A fast 'Optimization Controller' picks which mechanisms to use moment to moment, bundling them into what the paper calls Action Packages, while a slower 'Strategy Adapter' adjusts longer-term preferences and operator weights based on how past attempts played out. Across 32 benchmark groups and 14 competing configurations, OptiCom posted the best average rank and won outright in 23 of the 32 groups.
Most LLM-based optimizers still commit to one search mechanism for an entire run, even though the paper's own diagnostics show that what works early on often stops working once a search stalls or the budget runs low. Splitting decision-making into a fast layer and a slow layer is a practical answer to that mismatch, and it fits a broader pattern in AI research: when scaling up models gets expensive, labs increasingly try to squeeze more performance out of how inference time itself is spent.
Winning 23 of 32 benchmark groups is a solid result, not a sweep, and there is no word yet on released code or real-world deployment outside the paper's own test suite.