Researchers have introduced GroundAnything, a 4-billion-parameter grounding foundation model that challenges the sequential nature of autoregressive (AR) visual grounding. Traditional AR grounding models serialize spatial predictions, which introduces latency and imposes an artificial left-to-right causal order on output tokens. The new work instead treats grounding as visual evidence extraction, where objects, locations, and spatial relations are jointly constrained by the image and query without an intrinsic generation order.
To achieve this, GroundAnything adopts bidirectional diffusion with blockwise denoising, allowing spatial hypotheses to emerge in parallel and be refined iteratively. Its training pipeline combines grounding pretraining on public datasets and dedicated data engines, direct AR-to-diffusion conversion with joint AR and diffusion objectives, supervised fine-tuning, and GRPO-based reinforcement post-training.
Across 30 grounding benchmarks, the autoregressive variant GroundAnything-VLM reaches a new overall state of the art among similarly sized models at 72.42%, remaining competitive with GPT-6 Astra at 71.35%. With entropy-guided decoding, GroundAnything also surpasses the prior state of the art at this scale, averaging 61.75% versus 53.32% for the fast MTP-based LocateAnything model.
The team further explored decoding strategies, showing that an optional self-speculative mode achieves a 4.51× speedup over the AR counterpart with only a 0.74 percentage-point drop in COCO F1mIoU. Infrastructure experiments confirm that progressive inference optimizations translate parallel decoding into practical speedups, supporting efficient visual grounding in latency-sensitive real-world systems.
Code and weights are open source, and the project page is available for further details.