groundingpi/groundanything is a 4B Diffusion-based model that finds the position and extent of a target specified by text within an image. The repository description states that it bundles multiple grounding tasks into one through parallel decoding, and that optionally using self-speculative decoding yields 4.51 times the throughput of autoregressive inference on the COCO benchmark. The 2026-10-31 commit 7791c0e merges PR #15 (Skywalker-yqz/update-citation-1031) and touched only the citation instructions (BibTeX).
What changed is the format of the citation entry. README.md and README.zh-CN.md were modified with identical content, so the change totals 26 lines (10 added, 16 removed). The previous entry was in the form of @misc with eprint, archivePrefix, primaryClass, and url, and the new entry was changed to @article. The braces around the title disappeared, the authors are listed surname-first with the rest folded into 'others', and the arXiv identifier 2609.39600 was dropped from the url field and moved to the arXiv preprint notation in the journal field. The cite key was also shortened from yu2026groundanythingreconcilingparalleldecoding to yu2026groundanything. Since the LaTeX side reads the entry type and the journal field to typeset the citation information, this change appears to be matched to the README whose purpose was to cite the arXiv paper.
However, this commit touched only the citation blocks of the two documents; the code, model weights, and evaluation results did not change. The 4.51 times figure is a value presented by the repository description, not a result re-measured by this change, and it carries the conditions of COCO and autoregressive inference. Now that the cite key is shorter, references may break in literature that already uses that key, so it is safer for those writing papers to check the key before the correction. The README states that third-party materials keep their respective licenses as they are, but this commit did not touch such license statements either.