HyperEyes: Dual-Grained Efficiency-Aware Reinforcement Learning for Parallel Multimodal Search Agents

Researchers introduce HyperEyes, a parallel multimodal search agent that processes multiple entities concurrently rather than sequentially, achieving 9.9% higher accuracy with 5.3x fewer tool calls than comparable systems. The system combines visual grounding and retrieval into atomic actions and uses dual-level reinforcement learning to optimize both accuracy and inference efficiency, addressing a gap in existing multimodal AI benchmarks that ignore computational cost.

HyperEyes represents a meaningful advancement in multimodal AI agent design by fundamentally rethinking how search agents handle complex queries. Rather than the traditional sequential approach of issuing one tool call per entity, the system enables parallel dispatch of grounded queries within a single round, reducing redundant interaction overhead. This architectural shift directly addresses computational inefficiency in current multimodal systems.

The technical contribution centers on a Dual-Grained Efficiency-Aware Reinforcement Learning framework that operates at both trajectory and token levels. The TRACE reward mechanism progressively tightens efficiency constraints during training while preserving genuine multi-hop search capabilities. The On-Policy Distillation component solves the credit-assignment problem inherent in sparse outcome rewards by providing dense corrective signals from an external teacher model. The introduction of IMEB, a 300-instance human-curated benchmark jointly evaluating accuracy and efficiency, fills a critical evaluation gap in existing benchmarks that measure performance without considering computational costs.

For the broader AI industry, this work demonstrates that efficiency-aware training can coexist with accuracy improvements rather than requiring trade-offs. The 5.3x reduction in tool-call rounds has direct implications for deployment costs and latency in production systems. The parallel approach scales better for complex queries decomposing into independent sub-retrievals, a common pattern in real-world multimodal search scenarios. The methodology may influence how future foundation models are trained to balance capability with computational efficiency, particularly as multimodal agents become integral to AI applications.

• →Parallel query dispatching achieves 9.9% accuracy improvement while reducing tool calls by 5.3x versus sequential approaches
• →Dual-grained reinforcement learning framework optimizes both trajectory-level efficiency and token-level correction simultaneously
• →New IMEB benchmark addresses critical gap by jointly evaluating search accuracy and inference efficiency metrics
• →Efficiency-aware training demonstrates that computational cost reduction doesn't require sacrificing model capability
• →Architecture design combining visual grounding and retrieval into atomic actions enables concurrent multi-entity search