LocateAnything introduces Parallel Box Decoding (PBD) to overcome serialization bottlenecks in vision-language grounding. By predicting bounding boxes as atomic units in parallel, it achieves up to 2.5× higher throughput while improving localization quality (+3.8% F1 on LVIS, +1.8% on COCO).
论文原始摘要
Vision-language models (VLMs) commonly formulate visual grounding and detection as a coordinate-token generation problem, serializing each 2D box into multiple 1D tokens that are learned and decoded largely independently. This token-by-token decoding mismatches the coupled structure of box geometry and creates a practical inference bottleneck due to strictly sequential generation. We introduce LocateAnything, a unified generative grounding and detection framework based on Parallel Box Decoding (PBD). By decoding geometric elements such as bounding boxes and points as atomic units in a single step, LocateAnything preserves intra-box geometric coherence and unlocks substantial parallelism. We show that PBD improves both decoding throughput and localization accuracy. We further develop a scalable data engine and curate LocateAnything-Data, a large-scale dataset with more than 138 million training samples, substantially increasing data diversity for high-precision localization. Extensive evaluations show that LocateAnything advances the speed-accuracy frontier, achieving significantly higher decoding throughput while improving high-IoU localization quality across diverse benchmarks. The results highlight the complementary benefits of Parallel Box Decoding and large-scale training data in enabling efficient and precise unified visual grounding and detection.
Paper Collector 中文速览
提出并行框解码，实现高效统一视觉定位
方法概述
将边界框和点等几何元素作为原子单元单步解码，替代原有的逐token串行生成，保持几何一致性并提升并行度，配合大规模数据训练实现高效精准定位
核心贡献
提出并行框解码方法，构建1.38亿样本数据集，在速度和精度上均优于现有方法
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。
核心问题
How can vision-language models achieve both high-quality and low-latency grounding by overcoming the serialization bottleneck of next-token prediction approaches that convert 2D geometric objects into 1D token streams?
核心方法
The authors propose Parallel Box Decoding (PBD), which treats each bounding box as an atomic unit and predicts full coordinate sets in one parallel step rather than token-by-token. A dual-formulation training strategy jointly optimizes NTP sequences for causal reasoning and block-wise MTP for box-aligned predictions, with three inference modes (Fast, Slow, Hybrid) to balance throughput and accuracy. The framework is trained on LocateAnything-Data, a curated dataset with 12M images and 138M queries.
方法组件
- LocateAnything integrates Parallel Box Decoding into VLMs for visual detection and grounding.
- The framework comprises four main components: architecture, training strategy, inference mechanism, and dataset construction.
- The architecture uses Moon-ViT vision encoder and Qwen2.5 language decoder to preserve fine-grained spatial details.
- Coordinates are normalized to [0, 1000], discretized, and organized into blocks of constant length L=6.
- Four block types are defined: Semantic, Box, Negative, and End blocks for different functional purposes.
- Unoccupied positions within blocks are padded with tokens to ensure uniform tensor shapes for parallel decoding.
论点验证
The paper provides explicit dataset statistics in p_31: 'the dataset contains 12M unique images and 138M natural language queries.' Additional details are provided in Table 10 and Fig. 6 showing domain breakdowns.
The design choice is fully specified in p_5 and p_15. The paper describes treating bounding boxes as atomic units with block size L=6, predicting full coordinate sets in parallel.
The dual-formulation training strategy is fully specified in p_16-p_17, including the concatenated input sequence construction and the joint loss optimization.
The specific parameter L=6 is stated in p_15, with explanation that it accommodates 4 coordinate tokens plus 2 structural tokens.
The four block types are defined in p_15 with clear descriptions of each type's function.
The attention mask design is fully specified in p_18-p_21 and Fig. 4, with clear description of causal attention for NTP, block-causal pattern across MTP blocks, and bidirectional intra-block attention.
The three inference modes are described in detail in p_25-p_28, with clear specifications of each mode's behavior and use cases.
The specific threshold values (0.7 probability, 80 max-min difference) are stated in p_24 as the conditions for triggering ambiguity detection.
The specific sampling parameters (temperature=0.7, top-p=0.9) are stated in p_55.
The specific repetition penalty value (1.1) is stated in p_56.
The specific block size parameter (6) is stated in p_56.
The stream packing strategy is described in detail in p_51, including the target budget (36,864 tokens) and the three core mechanisms (Weighted Sampling, Best-Fit Buffering, Big-Rocks-First Seeding).
MagiAttention is described in p_52-p_53 as the framework for handling heterogeneous attention masks in the dual-formulation training.
The negative sample construction is described in p_61, with the rationale of mitigating hallucination behaviors.
The X-Y Corner Order is justified by ablation results in p_44 and Fig. 7 (left), showing it yields the highest F1-score among the four spatial sorting strategies tested.
The specific improvement numbers (+3.8% on LVIS, +1.8% on COCO) are stated in p_42, attributed to Table 1 results comparing to Rex-Omni with identical model size.
The throughput improvement is supported by Fig. 7 (right) showing 2× to 6× speedup, and Table 12 showing BPS values. The 'up to 2.5×' claim in p_8 is within the demonstrated range.
... 共 39 个论点
可复现性评估
较低可复现性 (0%)
缺失的复现细节
- Code is not available - no implementation details accessible
- Training data (LocateAnything-Data with 138M queries) is not available
- Core hyperparameters missing: learning rate, batch size, number of epochs, optimizer settings, weight decay, learning rate schedule
- Random seeds not specified for reproducibility
- Hardware specifications not provided (GPU type, number of GPUs, memory requirements)
- Training duration/computational cost not reported
- Detailed configurations referenced as being in 'supplementary materials' but these are not accessible
- Exact composition and sources of the 138M query training mixture not specified
- Base VLM pre-training details not provided (world-knowledge alignment phase)
- MLP projector architecture details not specified (dimensions, layers)
局限与证据边界
- Parallel decoding faces an inherent exploration-exploitation dilemma in highly complex scenes, with two failure patterns: Format Irregularity and Spatial Ambiguity.
- Format Irregularity occurs in complex scenes containing multiple instances across categories, where the model may struggle at category boundaries, hesitating between continuing to predict for the current class or transitioning to a new class.
- Spatial Ambiguity arises when objects are densely arranged in regular grids, such as rows or columns, where the MTP approach can blur spatial boundaries and output an intermediate coordinate situated between two objects.
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：2026-05-28T13:12:38+00:00 · 数据来源：Paper Collector