Hill climbing, often written hillclimb or hillclimbing, is a search and optimization method that repeatedly makes a small change to the current state, keeps the change if an evaluation score improves, and reverts it otherwise. It is a classic technique in mathematical optimization and artificial intelligence search. As a heuristic, it can get stuck in a local optimum rather than reaching the best overall solution, and convergence is not guaranteed.
In AI engineering, the term describes the practice of iteratively adjusting prompts, settings, or code and measuring each change against an eval score. To avoid overfitting to the evaluation data, the data is commonly split into train and test sets, and a change is kept only when scores improve on both.