CAMEO: Cooperative Agentic Multi-objective Evolution of Heuristics

We propose CAMEO (Cooperative Agentic Multi-objective Evolution of Heuristics), a cooperative multi-agent framework for heuristic discovery in multi-objective combinatorial optimization. Rather than treating heuristic generation as a sequence of isolated edits optimized by a single scalar reward, CAMEO formulates solver construction as a cooperative design process over two competing heuristic-level objectives: problem-level Pareto quality and runtime efficiency. Each candidate heuristic is represented as a point in a heuristic objective space, with its utility measured by the hypervolume improvement it contributes to the current archive of non-dominated heuristics. This turns heuristic discovery into a multi-objective search problem, aiming to construct a complementary set of heuristics rather than a single best solver. The resulting formulation is modular, Pareto-aware, and naturally compatible with language-model-based program editing. Our experimental results show that CAMEO can achieve a high-quality heuristic set across benchmarks.