iCORE framework audits LLM agents, boosting trajectory quality by 26.4%

As LLM-agent systems grow more autonomous, they self-organize to tackle complex tasks—but that emergent cooperation creates an auditability gap. Traditional logging captures messages, tool calls, and dependencies, yet fails to jointly represent what work remains, who is responsible for it, and what evidence justifies each state transition. Researchers Zuyuan Zhang, Hanqing Yang, Carlee Joe-Wong, and Tian Lan address this with iCORE (Integrated Cooperation-Obligation REpresentation), a unified encoding X=(G,Q,Π) that combines a cooperation graph (interactions), an obligation graph (work and assignments), and an audit map linking both to verifiable properties and evidence. This lets auditors certify two complementary properties: work soundness, meaning every active decision-relevant work assertion must have a finite justification; and agent-assignment stability, meaning no alternative agent can improve the declared contribution value for an obligation by more than ε.

Built as an instrumentation layer over workflows, iCORE provides local-to-global soundness guarantees and assignment-regret bounds. In experiments, the full coupled state exactly reconstructed soundness and assignment defects across two execution modes. Compared to passive full-state observation, iCORE-Audit delivered absolute trajectory-quality improvements of 11.5% in controlled runs and 26.4% with real LLMs, along with corresponding terminal-performance gains of 15.1% and 31.0%. This suggests that coupling collaborative interactions with obligation tracking not only catches hidden failures but also measurably improves output quality. For enterprises deploying multi-agent AI, iCORE offers a structured way to audit emergent behavior and ensure accountability—without trading off performance.