SplatLab: Autonomous Scientific Discovery for 3DGS Research

SplatLab. Autonomous Scientific Discovery for 3DGS Research. By Seemandhar Jain and one other, from the University of California, San Diego. Published on the twenty-first of October. Its code repository is worked out: about one hundred source files, with a readme and a licence.

Before anyone can extend a published paper on 3D Gaussian Splatting, someone has to reproduce it, and that takes weeks of expert engineering. Earlier autonomous research systems already read papers, draft hypotheses and write code. But they are held back by short fixed templates, low executability, and weak traceability, and none has been shown working end to end in a live computer-vision subfield with as much activity as this one. Those three weaknesses matter together. A template that is too narrow limits what can be tried, code that does not run means nothing gets measured, and without traceability a reported result cannot be checked against the experiment that supposedly produced it.

SplatLab is an autonomous research system built for 3D Gaussian Splatting. Give it a research prompt, a link to a target paper, or nothing at all, and it reproduces a published baseline, proposes an extension, trains and ablates on Mip-NeRF 360 and LLFF, writes the manuscript, and ties every number it reports to a logged training run. That is the whole research cycle, from reproduction to written paper, handled by one system, with each stage answering one of the weaknesses that held earlier systems back.

Five design choices set it apart. First, candidate ideas are built by mutating plugins on top of a backbone whose reproduction has been verified against the published baselines. Second, candidates compete in a cascade-weighted Elo tournament spread across eight islands, scored on PSNR measured on the GPU, with verbal-gradient feedback.

Third, a standing falsifier takes Elo away when a candidate fails, so surviving stress tests becomes part of selection. Fourth, every reported number is bound to a logged run, and a run cannot finish while any claim is still unbacked. Fifth, a taxonomy of 3DGS with seven categories wraps EVOLVE-BLOCK markers in scaffolds aligned to the domain, which can be carried across 3DGS repositories.

On runs targeting PSNR, the manuscripts SplatLab produces beat the targets they reproduced by between zero point one eight and three point seven eight decibels. The largest gain, three point seven eight decibels, is on rendering astrophotographic nebulae, a setting where no prior 3DGS method exists. The authors report the research cycle shrinking from months of a graduate student's time to days of mostly autonomous compute. Reproduction, which opened this story as the weeks-long bottleneck, becomes the starting point the system builds from rather than the obstacle in its way.

The code, every manuscript, the experiment ledgers and the cross-run paper store are available on the project page. The design leans hard on one principle: nothing gets written down unless a logged run stands behind it. The falsifier and the claim binding both serve that, and they are what separate a system that drafts papers from one whose numbers can be traced. Releasing the ledgers and manuscripts alongside the code lets a reader follow any reported figure back to the run it came from.