SDOF: Taming the Alignment Tax in Multi-Agent Orchestration with State Machine Constraints

Forum topic · 小凯 · 2026-05-19

Summary

SDOF is a framework that treats multi-agent LLM orchestration as a constrained state machine, addressing the lack of stage-constraint enforcement in frameworks like LangChain, LangGraph, and CrewAI. It uses two defensive layers implemented by three components: an Online-RLHF Specialized Intent Router trained via Generative Reward Modeling (GRPO), and a StateAwareDispatcher with GoalStage finite-automaton checks plus precondition/postcondition SkillRegistry validation for auditable execution control. Evaluated on a recruitment system backed by Beisen's iTalent platform (6000+ enterprises), 185 expert-curated scenarios triggered 1671 live API calls. The GSPO-aligned 7B Intent Router achieved 80.9% joint accuracy versus 48.9% for zero-shot GPT-4o on an FSM-constrained adversarial routing benchmark. End-to-end, SDOF reached 86.5% task completion (95% CI 80.8–90.7) and blocked all 22 injection and unauthorized HR-subset operations. In a broader message-level blocking audit, it achieved 100% precision, 88% recall, and expert agreement kappa=0.94. On 960 SGD-derived dialogues spanning 8 service domains, 201 stage-order conflicts emerged under FSM mapping, 41 in normal groupings. Paper: arXiv:2505.10891.

论文概要

Research area: ML Author: Zhantao Wang Published: 2025-05-15 arXiv: 2505.10891

Abstract (translated)

Multi-agent orchestration frameworks such as LangChain, LangGraph, and CrewAI route tasks through graph-based pipelines but do not enforce the stage constraints that govern real business processes. We present SDOF, a framework that treats multi-agent execution as a constrained state machine. SDOF operates through two primary defensive layers, implemented by three components: (1) an Online-RLHF Specialized Intent Router trained via Generative Reward Modeling (GRPO) and (2) a StateAwareDispatcher with GoalStage finite-automaton checks and precondition/postcondition SkillRegistry validation for auditable execution control. On a recruitment system backed by the Beisen iTalent platform (6000+ enterprises), 185 expert-curated scenarios trigger 1671 live API calls. Our GSPO-aligned 7B Intent Router exceeds zero-shot GPT-4o in joint accuracy (80.9% vs 48.9%) on the FSM-constrained adversarial routing benchmark. In end-to-end execution, SDOF achieves 86.5% task completion (95% CI 80.8–90.7) and blocks all 22 injection and unauthorized HR-subset operations. In a broader message-level blocking audit, SDOF achieves 100% precision and 88% recall, with expert agreement kappa=0.94. In a separate evaluation on 960 SGD-derived dialogues covering 8 service domains, 201 stage-order conflicts emerge under our FSM mapping, 41 of which appear in the normal grouping.

Key results

Routing: 7B GSPO-aligned router beats zero-shot GPT-4o — 80.9% vs 48.9% joint accuracy.
End-to-end execution: 86.5% task completion rate (95% CI: 80.8–90.7).
Safety: blocked all 22 injection and unauthorized HR-subset operations; 100% precision / 88% recall in message-level blocking audit (kappa = 0.94).
Scale of evaluation: 185 expert-curated scenarios, 1671 live API calls on Beisen iTalent (6000+ enterprises); 960 SGD-derived dialogues across 8 service domains.
Findings: 201 stage-order conflicts surfaced in the SGD evaluation, including 41 in normal groupings.

Components

1. Online-RLHF Specialized Intent Router — trained via Generative Reward Modeling (GRPO). 2. StateAwareDispatcher — GoalStage finite-automaton checks plus precondition/postcondition SkillRegistry validation for auditable execution control.

---

*Auto-collected on 2026-05-19*