LocateAnything is a unified generative visual grounding and detection framework designed to overcome the autoregressive bottlenecks prevalent in Vision-Language Models (VLMs). Traditional VLMs typically formulate visual grounding by serializing 2D bounding boxes into multiple 1D coordinate tokens, which are then decoded largely independently and sequentially. This token-by-token decoding method inherently mismatches the coupled structure of box geometry and creates a significant inference bottleneck due to its strict sequential nature.

The core innovation of LocateAnything is its Parallel Box Decoding (PBD) methodology. Instead of segmenting a bounding box into individual coordinate tokens, PBD treats each bounding box (defined by $(x_1, y_1, x_2, y_2)$) or a point as an atomic unit of constant length. The model then predicts the full coordinate set for multiple such atomic units simultaneously in a single, parallel step. This approach fundamentally preserves intra-box geometric coherence and unlocks substantial parallelism, leading to improved decoding throughput and enhanced localization accuracy.

The Architecture of LocateAnything is built upon a Moon-ViT vision encoder, which extracts visual tokens at native resolution to preserve fine-grained spatial details crucial for high-precision localization. A Qwen2.5 language decoder processes the textual queries. These two components are bridged by a Multi-Layer Perceptron (MLP) projector, which directly converts visual tokens into a sequence of "box-aligned block-level predictions." This means the output is structured to inherently predict complete box representations rather than individual coordinate components.

LocateAnything offers Flexible Inference Modes to balance speed and accuracy:

1. Fast Mode (MTP - Massively Parallel Prediction): This is the primary PBD mechanism where the model predicts full boxes in parallel for maximum throughput. It is suitable for latency- and compute-constrained scenarios, such as on-device robotics or embodied agents.
2. Slow Mode (NTP - Next Token Prediction): In this mode, the model decodes coordinate tokens autoregressively, similar to traditional VLM approaches. While slower, it offers maximum stability and precision, making it appropriate for high-precision labeling, dataset curation, and accuracy-oriented offline evaluation. The paper's ablation studies indicate that a box-aligned formulation, even in Slow Mode, provides stronger supervision than 1D serialization.
3. Hybrid Mode: This is the default operational mode, designed to achieve a robust balance. It primarily utilizes Fast Mode for its speed benefits but dynamically falls back to Slow Mode when the model detects "Format Irregularity" (malformed syntax at category boundaries) or "Spatial Ambiguity" (intermediate coordinates between densely arranged objects). This Corrected NTP Re-decoding mechanism ensures robust predictions by discarding compromised blocks, reverting to the last verified prefix, generating tokens for the problematic block autoregressively using NTP, and then switching back to MTP. This strategy preserves most of the speed gains while maintaining output integrity.

To train this highly capable model, the authors curated LocateAnything-Data, a large-scale dataset comprising over 138 million training samples and 785 million boxes across 12 million unique images. This dataset significantly increases data diversity for high-precision localization and encompasses a wide array of query types:

• General Object Detection (66.9% of queries): Provides essential bounding box supervision.
• GUI Element Grounding (16.5%): Supports embodied agents and graphical user interface navigation.
• Referring Comprehension (7.3%): Links complex natural language intents to specific spatial regions.
• Text Localization (OCR) (3.6%): Perceives and grounds textual information.
• Layout Grounding (3.5%): Enriches structural reasoning for document and scene understanding.
• Point-Based Localization (2.2%): Refines spatial precision for fine-grained predictions.

Main Results demonstrate that LocateAnything significantly advances the speed-accuracy frontier. It achieves 12.7 Boxes Per Second (BPS) in Hybrid Mode on a single NVIDIA H100 GPU, which is over 10 times faster than textual-based Qwen3-VL (1.1 BPS) and 2.5 times faster than quantized Rex-Omni (5.0 BPS). LocateAnything shows substantial improvements in high-IoU localization quality across diverse benchmarks. For instance, it improves mean F1 by +3.8% on LVIS and +1.8% on COCO compared to Rex-Omni, with particularly strong gains at high IoU thresholds (e.g., 31.1 vs. 20.7 at IoU=0.95 on LVIS). It also excels in dense object detection (Dense200, VisDrone), precise open-world localization (ScreenSpot-Pro for GUI grounding), layout grounding (DocLayNet, M6Doc), OCR (TotalText), and referring expression comprehension (HumanRef, RefCOCOg).

Ablation studies confirm the efficacy of PBD:

• The PBD (Slow Mode) formulation achieves the highest F1 of 52.1 on COCO, proving that box-aligned supervision is superior to 1D serialization.
• PBD significantly outpaces structure-agnostic MTP methods (16.9 BPS vs. 5.5 BPS for SDLM-B6) while improving F1.
• Joint training of both Fast and Slow Modes pushes the Slow Mode upper bound from 50.1 to 52.1 F1, and the Hybrid Mode effectively balances speed and accuracy, achieving 51.6 F1 at 13.2 BPS.
• Regarding box ordering, X-Y Corner Order sorting yields the highest F1-score.
• PBD achieves a 2x to 6x speedup over NTP methods as the number of target boxes increases from 20 to 300, scaling throughput from 12 BPS to approximately 25 BPS in dense scenes.