SDOF Framework Reduces Alignment Tax in Multi-Agent Orchestration

The SDOF (State-Constrained Dispatch Orchestration Framework) introduces a solution to the alignment tax in multi-agent orchestration by conceptualizing execution as a constrained state machine. This framework, created by researchers, incorporates two protective layers: an Online-RLHF Specialized Intent Router, which is trained using Generative Reward Modeling (GRPO), and a StateAwareDispatcher that performs GoalStage finite-automaton checks alongside SkillRegistry validation for preconditions and postconditions. Implemented on the Beisen iTalent platform, which supports over 6,000 companies, SDOF utilized 185 expert-curated scenarios, resulting in 1,671 live API calls. The GSPO-aligned 7B Intent Router achieved a joint accuracy of 80.9% on a constrained adversarial routing benchmark, surpassing zero-shot GPT-4o's 48.9%. This framework seeks to enforce stage constraints in actual business processes, enhancing reliability and auditability in multi-agent systems.