XPENG's GroundAnything pairs diffusion parallel decoding with precise grounding, 4.51x faster than AR
XPENG-AI · hf · 2026-10-13
From the same XPENG-led team, GroundAnything is a second grounding foundation model that combines fast parallel decoding with precise localization:
- Idea: treat grounding as visual evidence extraction — spatial predictions carry no inherent left-to-right order, so bidirectional diffusion with blockwise denoising lets spatial hypotheses emerge in parallel and be jointly refined.
- Training: AR-to-diffusion conversion with joint AR and diffusion objectives, plus SFT and GRPO-based RL post-training.
- Results: GroundAnything-VLM (the AR variant) sets a same-size SOTA of 72.42% across 30 benchmarks, competitive with GPT-6 Astra (71.35%); with entropy-guided decoding it averages 61.75% vs 53.32% for the prior fast SOTA LocateAnything.
- Speed: an optional self-speculative mode is 4.51x faster than the AR counterpart with only a 0.74 pp drop in COCO F1mIoU, reaching up to 223.79 TPS.
- Why it matters: efficient visual grounding for latency-sensitive systems like robots, driving, and GUI agents.
More from Research
- AgentSWE benchmark: no coding agent comes close to delivering the agents it's asked to build — Jiahao Wang · 2026-10-13
- Schema agent hits ~99% RHAE on ARC-AGI-3 public set, full paper and code released — wzenus · 2026-10-13
- World model research: more data won't suffice—squeeze more from it; even intuitive physics is still out of reach — mangahomanga · 2026-10-13
- Indie dev builds own serving library to train SVG model, finds evolution loses to SGD — cephaloform · 2026-10-13
- Harvard & NTT Research: cooperating AI agent populations hit an Allee-effect takeoff threshold — Hidenori8Tanaka · 2026-10-13
- RT-Super trains tumor segmentation from radiology reports, lifting Dice from 16 to 30 without masks — maier_ak · 2026-10-13