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

2026-05-07

AI Summary

Existing multimodal search agents suffer from inefficiency due to sequential processing of target entities, resulting in redundant query decomposition and excessive interaction rounds. This work proposes HyperEyes, which unifies visual grounding and retrieval into atomic actions, enabling concurrent single-round search for multiple entities with inference efficiency as the central optimization objective. We introduce a novel dual-granularity, efficiency-aware reinforcement learning framework: at the macro level, a dynamic TRACE trajectory reward mechanism progressively tightens reference constraints; at the micro level, online distillation supplies fine-grained corrective signals. Additionally, we present IMEB, the first human evaluation benchmark that jointly assesses accuracy and efficiency. Experiments show that HyperEyes-30B achieves a 9.9% absolute accuracy gain over the strongest open-source agent across six benchmarks while reducing average tool-call rounds by 5.3×.

Abstract

Existing multimodal search agents process target entities sequentially, issuing one tool call per entity and accumulating redundant interaction rounds whenever a query decomposes into independent sub-retrievals. We argue that effective multimodal agents should search wider rather than longer: dispatching multiple grounded queries concurrently within a round. To this end, we present HyperEyes, a parallel multimodal search agent that fuses visual grounding and retrieval into a single atomic action, enabling concurrent search across multiple entities while treating inference efficiency as a first-class training objective. HyperEyes is trained in two stages. For cold-start supervision, we develop a Parallel-Amenable Data Synthesis Pipeline covering visual multi-entity and textual multi-constraint queries, curating efficiency-oriented trajectories via Progressive Rejection Sampling. Building on this, our central contribution, a Dual-Grained Efficiency-Aware Reinforcement Learning framework, operates at two levels. At the macro level, we propose TRACE (Tool-use Reference-Adaptive Cost Efficiency), a trajectory-level reward whose reference is monotonically tightened during training to suppress superfluous tool calls without restricting genuine multi-hop search. At the micro level, we adapt On-Policy Distillation to inject dense token-level corrective signals from an external teacher on failed rollouts, mitigating the credit-assignment deficiency of sparse outcome rewards. Since existing benchmarks evaluate accuracy as the sole metric, omitting inference cost, we introduce IMEB, a human-curated benchmark of 300 instances that jointly evaluates search capability and efficiency. Across six benchmarks, HyperEyes-30B surpasses the strongest comparable open-source agent by 9.9% in accuracy with 5.3x fewer tool-call rounds on average.

Problem

• multimodal search
• parallel agent
• tool-call efficiency
• concurrent retrieval
• inference cost

Innovation

• Parallel Multimodal Search
• Efficiency-Aware Reinforcement Learning
• Dual-Grained Reward
• TRACE
• On-Policy Distillation