Large language models are increasingly being used to discover heuristics through iterative program search, a promising route to automated algorithm design. The approach typically evaluates candidate programs by their average performance across a set of tasks.
However, the abstract warns that candidates selected purely on average performance may degrade—likely in scenarios where worst-case outcomes matter. This suggests that average-based selection can produce heuristics that are not robust to distribution shifts or difficult instances.
The paper introduces CVaR-RMD, which applies CVaR selection pressure during the search. By focusing on the conditional value at risk rather than the mean, the method aims to steer discovery toward heuristics that maintain performance under adverse conditions. The work highlights a growing concern in LLM-driven algorithm discovery: optimizing for the average may come at the cost of reliability.