SKIMIX Framework Uses Skill Mixture for Multi-Agent Scaling

SKIMIX employs embedding-based skill retrieval and submodular anti-dilution routing to manage multi-agent collaboration via iterative refinement. Benchmarks show significant gains in open-ended mathematical reasoning, though agent-count scaling is non-monotonic and performance on multiple-choice tasks remains limited.