SplatLab：面向3DGS研究的自主科学发现

arXiv 2026 73.2 method

TLDR

SplatLab autonomously reproduces 3DGS papers, proposes extensions, runs experiments, and writes traceable manuscripts, beating baselines by up to +3.78 dB PSNR.

评分理由

Strong end-to-end demonstration with verified reproduction, traceable claims, and tournament-based selection; however, scope is narrow (3DGS/PSNR) and evidence of broader scientific discovery or human-level validation is not shown.

Read-first 评分解释

综合优先阅读分 73.2，由主题、引用、图谱、方法、可复现性和近期性等信号加权得到。原始总分保留为 89。

近期性 8%

100

使用温和的时间衰减，让近期论文更容易浮现，同时保留较早基础工作的价值。年份：2026

方法质量 25%

检查可见的摘要与分析字段，寻找实验、数据集、基线、指标和局限性等方法证据。命中信号：基线、评估、实验、指标

主题相关性 42%

74.2

使用现有 LLM 关键词相关性评分，并归一化到 0-100。关键词：AI scientist、automated scientific discovery、autonomous research agent、automated research、literature review agent、survey generation、automated experimentation、experiment design agent、AI for scientific research、paper writing agent、research automation、scientific discovery agent

可复现性 25%

检查链接和可见文本中的论文、代码、数据集、工件与仓库信号。论文：有；代码：无；数据：无；命中信号：代码、复现

研究版图角色

前沿论文方法锚点

排序敏感性

稳定性：volatile；排名波动范围：27。

关键词评分

automated scientific discovery

autonomous research agent

automated experimentation

AI for scientific research

research automation

scientific discovery agent

AI scientist

automated research

experiment design agent

paper writing agent

literature review agent

survey generation

深度分析

创新点

End-to-end autonomous research system for 3D Gaussian Splatting that reproduces baselines, proposes extensions, trains and ablates, writes manuscripts, and binds reported numbers to logged training runs.
Candidates mutate plugins from a verified-reproduction backbone matched to published baselines.
A cascade-weighted Elo tournament across eight islands scores candidates on measured GPU PSNR with verbal-gradient feedback.
A standing falsifier subtracts Elo on failure, integrating stress-test survival into selection.
Every reported number is bound to a logged training run, and the run cannot terminate while any claim is unbacked.
A seven-category 3DGS taxonomy wraps EVOLVE-BLOCK markers in domain-aligned scaffolds transferable across 3DGS repositories.

方法

SplatLab takes a research prompt, a target-paper URL, or no input; it reproduces a published 3DGS baseline, proposes an extension, trains and ablates across Mip-NeRF 360 and LLFF, writes a manuscript, and binds each reported number to a logged training run. Candidate methods are generated by mutating plugins on a verified-reproduction backbone and selected through a cascade-weighted Elo tournament across eight islands using measured GPU PSNR, verbal-gradient feedback, and a standing falsifier. A seven-category 3DGS taxonomy and EVOLVE-BLOCK markers provide domain-aligned scaffolds for transferability.

关键结果

On PSNR-targeted runs, SplatLab's manuscripts beat their reproduced targets by +0.18 to +3.78 dB, including +3.78 dB on astrophotographic nebula rendering, for which no prior 3DGS method exists. The research cycle compresses from months of graduate-student time to days of mostly autonomous compute.

局限性

The abstract does not state explicit limitations of SplatLab itself.
Reported gains are for PSNR-targeted runs; no other evaluation metrics or objectives are described in the abstract.

资源

论文