3D Gaussian Splatting research without hand-written code: SplatLab runs from a one-line prompt to a manuscript, each beating its reproduced baseline

Reproducing and extending a 3DGS paper no longer depends entirely on graduate-student engineering; a research prompt, a target-paper URL, or no input at all starts a run that goes from reproduction to manuscript. Prior autonomous-research systems edited short LaTeX templates, ranked free-text hypotheses over brief training, or built on LLM-written reimplementations not externally checked against the target; across 20 ICML 2024 Spotlight and Oral papers, the strongest evaluated agent reached 21.0% average replication. Across 16 runs SplatLab trained 198 experiment branches, yielding five verified positives, five marginal results with qualified claims, five non-PSNR contributions and one recorded failure; runs used a 24–48-hour wall-clock budget, averaging 113M API tokens over roughly 1,000–1,500 LLM calls and 2–36 GPU-hours.

Candidates do not write code from scratch; each mutates a plugin that already compiles, trains and matches its target's published headline, so gains can be attributed to the proposed mechanism. Without a verified reproduction an improvement claim has no fixed external reference, reported metrics may not trace to a measurement on disk, and long-running agents drift from their objective as context fills. In the ablation, removing the verified-plugin contract left 9 of 16 winners non-reproducible versus 0 of 16 with it; across the eight baseline pairings on Mip-NeRF 360 and LLFF, the generated methods exceed their baselines on PSNR, with the largest gain on the nebula scene where 3DGS had not previously been applied.

Idea ranking is decided by PSNR measured on real GPUs rather than by model-judged proposal text. Earlier Elo-style ranking scored free-text hypotheses and updated ratings uniformly rather than scaling by evaluation depth. Candidates pass through smoke, probe, multi-scene and full-benchmark stages with cascade-weighted Elo updates; in the ablation, uniform weights picked the smoke-test winner, which placed fifth on multi-scene, while the full pipeline picked the multi-scene leader.

Every number in a manuscript must correspond to a logged training run, and a failed audit aborts the run as a research log. Earlier automated writing pipelines allowed drafts to carry numbers with no backing measurement. In the ablation, without the claim-backed audit gate all 16 manuscripts contained numbers that did not trace to any log on disk, versus 0 of 16 with the full pipeline.

The demonstrated scope is 3D Gaussian Splatting: standard benchmarks are Mip-NeRF 360 and LLFF at 30k training iterations on NVIDIA A6000 GPUs, plus nebula, bubble and diamond scenes captured for this work. It is aimed at researchers following autonomous-research systems and neural rendering; the source lists cross-domain transfer and taxonomy adaptation as steps needing expert effort.

The source states the experiments demonstrate system capability under end-to-end operation rather than equalized search budgets; the taxonomy ablation compares only the nebula target, on which the authors conclude domain grounding improves search efficiency rather than being strictly necessary for execution. Human reviewers scored the top five generated papers at 3.5 overall, borderline, versus 4.1 for CVPR/NeurIPS 2025 published papers, while the multi-agent reviewer gave 3.6 versus 3.5.