Before anyone can improve on a published gaussian splatting paper, they usually spend weeks reproducing it. The training schedule and a long list of small implementation details have to match the original before a new idea can be fairly measured against it.
SplatLab, from Seemandhar Jain and Manmohan Chandraker at UC San Diego, hands that entire cycle to a team of software agents. Given a research prompt, a link to a target paper, or no input at all, SplatLab reproduces a published baseline, proposes an extension, trains and tests it, and writes the manuscript. The paper has been accepted at NeurIPS 2026, where the conference listing titles it "Who Wrote This Paper? Autonomous Scientific Discovery for 3DGS Research"
Every candidate idea starts from a reproduction already checked against the original paper's published numbers, built with the group's earlier SPLATIFY system. Candidates then compete in a tournament spread across eight separate pools, ranked on image quality measured from real training runs. A standing critic attacks each survivor with difficult scenes, fewer input views and shifted settings, and docks rating points from any that fail.
SplatLab also links every number in a draft to a logged training run, and a run cannot finish while any claim lacks one. In the authors' ablations, switching that check off left unbacked numbers in all 16 manuscripts.
Across two standard benchmark sets and three new test scenes, SplatLab's manuscripts beat their reproduced targets by 0.18 to 3.78 dB in PSNR. The largest gain came on a nebula scene captured for the paper, an astrophotography case the authors say no prior 3DGS method has addressed. Each run had a 24 to 48 hour wall clock budget, with estimated API costs of $150 to $250.
Three generated manuscripts were accepted as posters at a CVPR 2026 workshop. A blind human reviewer scored five SplatLab papers higher on novelty than five published papers from CVPR and NeurIPS 2025, but lower on soundness and presentation, placing them at borderline against a clear accept for the human work. The authors recommend that papers produced this way be labeled and ship with their experiment logs, and they discourage volume submissions.
SplatLab's code is available on GitHub under the MIT license, and the paper is available now on arXiv.
