EARS framework boosts multi-agent reliability by 10.4 percentage points

In large-scale multi-agent systems, sub-agents built on smaller fine-tuned models often over-answer ambiguous or misrouted requests, producing hallucinated outputs instead of actionable feedback. Researchers from an unnamed institution propose EARS, a production framework that reframes abstention as an inter-agent communication protocol: sub-agents expose failure states (e.g., underspecification, capability mismatch) with structured rationales, enabling the coordinator to clarify, reroute, or fallback.

EARS uses an ensemble of calibrated LLM-as-a-Judge models to curate human-agent interaction data under a taxonomy of failure modes. The sub-agents are fine-tuned to detect these conditions and return rationales. Evaluated on a production e-commerce assistant for enterprise BI, the system boosted the overall pass rate from 68.5% to 78.9%, demonstrating that explanatory abstention significantly improves reliability without sacrificing speed.