TL;DR
HyperEyes presents a parallel multimodal search agent using Unified Grounded Search and Dual-Grained Efficiency-Aware RL. HyperEyes-30B surpasses the strongest open-source agent by 9.9% accuracy while requiring 5.

论文原始摘要

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.

Paper Collector 中文速览

HyperEyes通过并行多模态搜索和双粒度效率感知强化学习提升搜索效率

方法概述

两阶段训练：先通过渐进拒绝采样合成并行友好数据；再用双粒度强化学习优化，宏观层采用TRACE轨迹级奖励自适应抑制冗余调用，微观层用On-Policy蒸馏注入token级校正信号

核心贡献

提出并行多模态搜索框架，融合视觉定位与检索为原子操作，设计双粒度效率感知强化学习（TRACE奖励+On-Policy蒸馏），构建IMEB效率评测基准

原始来源与核验范围

核验仅覆盖论文正文中的主张、证据片段和参考文献关系。NGJOO 未独立运行作者代码、重做实验或验证真实部署效果；页面中的可复现性分数是文档完整度评估，不是复现实验结果。

43
已证实
0
证据不足
3
无法验证
N/A
可复现性
置信度
82%

核心问题

How can multimodal search agents be optimized to perform parallel tool invocations instead of sequential ones, reducing interaction redundancy while maintaining or improving accuracy?

核心方法

HyperEyes uses a Unified Grounded Search action space that fuses visual grounding and retrieval into a single atomic action for parallel execution. The training pipeline combines Supervised Fine-Tuning on efficiency-filtered trajectories with a Dual-Grained RL framework featuring TRACE for trajectory-level efficiency guidance and On-Policy Distillation for token-level supervision.

论点验证

已证实 (85%) We propose HyperEyes, a parallel multimodal search agent designed around the principle of 'search wider, not longer.'
已证实 (85%) HyperEyes operates on a Unified Grounded Search (UGS) action space that fuses visual grounding and retrieval into a single atomic action, extending text-level parallelism to the visual modality.
已证实 (80%) We pair this architecture with a Dual-Grained Efficiency-Aware reinforcement learning (RL) framework that treats efficiency as a primary optimization objective.
已证实 (80%) At the macro level, it features TRACE, a trajectory-level reference that dynamically tightens during training to guide the policy toward optimal efficiency.
已证实 (80%) At the micro level, it introduces On-Policy Distillation (OPD), which resolves ambiguous credit assignment by providing dense per-token supervision from an expert teacher on failed rollouts.
已证实 (85%) We support this training paradigm with a Parallel-Amenable Data Synthesis Pipeline, which utilizes Progressive Rejection Sampling to curate high-quality, efficiency-oriented cold-start trajectories.
已证实 (80%) We introduce IMEB, the first human-curated benchmark to jointly evaluate answer accuracy and search efficiency, establishing operational efficiency as a first-class metric in multi-entity visual scenarios.
已证实 (85%) HyperEyes-30B establishes state-of-the-art results. It Pareto-dominates existing models, surpassing the strongest open-source agent by 9.9% in accuracy while requiring 5.3× fewer tool-call rounds on average.
Specific quantitative results stated in p_48 and verifiable from Table 2. HyperEyes-30B achieves 9.9% higher accuracy than VDR with 5.3× fewer tool calls. These are concrete numbers from controlled comparisons against baselines.
已证实 (80%) We address this with Unified Grounded Search (UGS), reformulating visual grounding from a prerequisite step into a parameter of the retrieval action.
UGS design is described in p_16, explaining how visual grounding becomes a parameter of the retrieval action rather than a prerequisite step. The controlled experiments in Figure 4 validate this design choice.
已证实 (80%) By simultaneously predicting bounding boxes for all target entities, UGS allows the policy to dispatch parallel search queries across modalities within a single turn.
The parallel dispatch capability is described in p_16 and demonstrated in controlled experiments (Figure 4, p_95) showing UGS achieves fewer tool-call turns (2.04) compared to sequential paradigms (2.7/3.07).
已证实 (75%) We compile a rich foundation of 246,000 multi-hop reasoning and visual recognition queries from existing public benchmarks and internal human annotations.
The number 246,000 is stated in p_19. However, the breakdown between public benchmarks and internal annotations is not detailed, and the exact sources are not comprehensively listed.
已证实 (80%) We supplement this pool with 25,000 novel synthetic queries across two bespoke pipelines.
The number 25,000 synthetic queries is stated in p_19, with detailed breakdown in p_20-21 (20,000 visual + 5,000 textual). The arithmetic is consistent: 246,000 + 25,000 = 271,000.
已证实 (80%) This pipeline yields 20,000 visual multi-entity QA pairs.
Specific number stated in p_20 and confirmed in p_69. The visual multi-entity synthesis pipeline is described in detail with source datasets and methodology.
已证实 (80%) This textual pipeline contributes an additional 5,000 complex queries.
Specific number stated in p_21 and confirmed in p_69. The textual multi-constraint synthesis pipeline is described with Wikidata source and constraint-chain construction.
已证实 (75%) We apply a unified filter across all task sources, systematically discarding any QA pair that Qwen3-VL-235B successfully resolves without external tool access, thereby finalizing our foundational pool of 271,000 genuinely tool-dependent tasks.
The filtering process is described in p_22 using Qwen3-VL-235B. The final number 271,000 matches the sum of 246,000 + 25,000. However, the exact filtering rate/rejection statistics are not provided.
已证实 (80%) To obtain a clean, efficiency-oriented training signal, we propose Progressive Rejection Sampling (PRS).
PRS is described in p_23 with Algorithm 1 provided. The mechanism of ascending budget schedule and shortest-trajectory retention is clearly specified.
已证实 (80%) Through this cascade of sampling and quality filtering, the initial pool of 271,000 tasks is distilled to 30,000 high-fidelity trajectories.
Specific numbers stated in p_24: 271,000 → 30,000 trajectories. The distillation ratio (~11%) is concrete. Table 4 ablation validates the quality filtering approach.
已证实 (80%) We isolate 6,056 and 9,337 queries for the 30B and 235B models, respectively, where the initial model fails to find an answer under the tightest pass@1 setting but successfully resolves the task under relaxed pass@5 constraints.
Specific numbers stated in p_25 and p_77: 6,056 for 30B and 9,337 for 235B. The pass@1 vs pass@5 selection criteria is clearly described.
已证实 (75%) We employ a two-stage agentic training paradigm. We first fine-tune the model on the curated demonstration corpus to instill basic parallel retrieval behaviors. Subsequently, we apply a Dual-Grained Efficiency-Aware RL framework to optimize search efficiency and token-level credit assignment.
The two-stage paradigm is described in p_26. However, while the stages are specified, the paper does not provide a detailed comparison of SFT-only vs SFT+RL to isolate the contribution of each stage.
已证实 (70%) The Supervised Fine-Tuning (SFT) phase optimizes the base MLLM via next-token prediction on the curated trajectory corpus.
The SFT phase is described in p_27 as next-token prediction on trajectory corpus. This is standard practice but not extensively validated in isolation.

... 共 46 个论点

可复现性评估

较低可复现性 (0%)

缺失的复现细节

局限与证据边界

本分析由 PDF 阅读助手 自动生成，仅供参考，不构成学术评审意见。验证结论和可复现性评估基于论文文本自动分析，可能存在偏差。原始论文请参阅 arXiv。

分析时间：2026-05-12T13:36:01+00:00 · 数据来源：Paper Collector