EARS framework boosts multi-agent reliability by teaching sub-agents to abstain
EARS is a production-oriented framework that fine-tunes sub-agents in large-scale multi-agent systems to detect their own failure conditions and return structured, actionable rationales to a coordinator instead of hallucinating responses.
Score breakdown
EARS converts sub-agent silence into structured, coordinator-actionable failure signals, directly raising the production response pass rate from 68.5% to 78.9% in a real enterprise deployment.
- 01EARS stands for Explanatory Abstention for Reliable Sub-Agent Modeling.
- 02The framework targets centralized MAS where a coordinator delegates requests to lightweight, domain-specialized sub-agents.
- 03Sub-agents built on smaller fine-tuned models often over-answer ambiguous or misrouted requests, producing hallucinated outputs.
In large-scale enterprise deployments, centralized multi-agent systems (MAS) typically rely on a coordinator that delegates user requests to lightweight, domain-specialized sub-agents. While this architecture improves modularity, scalability, and cost efficiency, its reliability hinges on sub-agents accurately calibrating their responses to their own capability limits. Sub-agents built on smaller fine-tuned models frequently fail at this calibration, over-answering ambiguous, underspecified, misrouted, or unsupported requests and generating hallucinated outputs rather than useful feedback.
EARS (Explanatory Abstention for Reliable Sub-Agent Modeling) reframes sub-agent abstention as an inter-agent communication protocol.
EARS (Explanatory Abstention for Reliable Sub-Agent Modeling) reframes sub-agent abstention as an inter-agent communication protocol. Instead of simply declining a request, a sub-agent exposes an actionable failure state — including a structured rationale — back to the coordinator, which can then clarify, reroute, or trigger a fallback. To produce training data for this behavior, EARS curates human-agent interaction data using an ensemble of calibrated LLM-as-a-Judge models, generating structured abstention labels and rationales organized under a taxonomy of sub-agent failure modes. Sub-agents are then fine-tuned on this data to detect failure conditions and return coordinator-actionable rationales.
The framework was evaluated in a large-scale production e-commerce assistant supporting enterprise business intelligence workflows. EARS improved the overall response pass rate from 68.5% to 78.9%, demonstrating that sub-agent-side explanatory abstention meaningfully improves MAS reliability at production scale.
Key facts
- 01EARS stands for Explanatory Abstention for Reliable Sub-Agent Modeling.
- 02The framework targets centralized MAS where a coordinator delegates requests to lightweight, domain-specialized sub-agents.
- 03Sub-agents built on smaller fine-tuned models often over-answer ambiguous or misrouted requests, producing hallucinated outputs.
- 04EARS reframes abstention as an inter-agent communication protocol: sub-agents return structured failure states with rationales rather than simply refusing.
- 05Training data is curated using an ensemble of calibrated LLM-as-a-Judge models that produce structured abstention labels and rationales under a taxonomy of failure modes.
- 06EARS was evaluated on a large-scale production e-commerce assistant for enterprise business intelligence workflows.
- 07The framework improved the overall response pass rate from 68.5% to 78.9%.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 18, 2026 · 10:40 UTC. How this works →