Paper 2605.07177
HyperEyes: Dual-Grained Efficiency-Aware Reinforcement Learning for Parallel Multimodal Search Agents
- Published
- May 2026
- Research lab
- Independent
- Citations
- 4
- GitHub
- 71 stars
01 In brief
Summary
HyperEyes is a parallel multimodal search agent that addresses inefficiencies in sequential search agents for multi-entity queries.
It introduces a Unified Grounded Search (UGS) action space, fusing visual grounding and retrieval into a single atomic action to enable concurrent search.
Training uses a two-stage approach: a Parallel-Amenable Data Synthesis Pipeline with Progressive Rejection Sampling for cold-start supervision, and a Dual-Grained Efficiency-Aware Reinforcement Learning framework.
This framework includes TRACE, a trajectory-level reward with a monotonically tightening reference to suppress redundant tool calls, and On-Policy Distillation (OPD) for token-level correction on failed rollouts.
The authors also introduce IMEB, a human-curated benchmark of 300 multi-entity visual instances for joint evaluation of accuracy and efficiency.
Across six benchmarks, HyperEyes-30B surpasses the strongest open-source agent of comparable scale by 9.9% in accuracy with 5.3× fewer tool-call rounds on average.
HyperEyes-235B approaches Gemini-3.1-Pro's accuracy.
Ablations show that quality filtering, adaptive efficiency references, and an efficiency-aligned teacher are crucial for performance.
02 From the paper
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.