🧪 SplatLab

SplatLab: Autonomous Scientific Discovery for 3DGS Research

🧪 Turn a one-line prompt (or nothing at all) into a reproduced baseline, a new method, real GPU experiments, and a compiled paper. 🧪

Seemandhar Jain · Manmohan Chandraker

University of California, San Diego

🏆 Accepted at NeurIPS 2026 🧪

---

🚀 The research cycle (read, reproduce, extend, train, ablate, write) compressed from months of graduate-student time to days of mostly autonomous compute.

📰 News

• 2026-09 — SplatLab is accepted at NeurIPS 2026! 🎉
• 2026-10 — Project page, paper PDF, interactive demo, and all 16 SplatLab-written papers are online.
• 2026-06 — 3 SplatLab-written papers accepted by human reviewers at a CVPR 2026 workshop.

Overview

SplatLab is an autonomous research system for 3D Gaussian Splatting. You give it a research prompt, a target-paper URL, or no input at all. It reproduces a published baseline, proposes an extension, trains and ablates it across Mip-NeRF 360 and LLFF on real GPUs, writes the manuscript, and ties every reported number to a logged training run.

It runs natively on Claude Code. One /research slash command turns the main session into a lab PI that runs for 24–48 h, dispatching 32 specialised subagents and coordinating them through a crash-safe ledger.json. An optional zero-dependency web dashboard can launch, watch, stop, and resume runs from the browser.

Five design choices set it apart:

	Design choice	What it buys
🧬	Verified-reproduction backbone	No experiment spawns until the target is reproduced within 0.2 dB on our GPUs; every candidate mutates a plugin already matched to its published baseline.
🏆	Grounded Elo tournament	A cascade-weighted Elo tournament across eight islands ranks candidates on measured ns-train PSNR — never model-judged scores — with verbal-gradient feedback to the ideator.
⚔️	In-loop falsification	Every survivor faces edge-scene, sparse-view, and load-bearing-hyperparameter attacks; each failure subtracts Elo, so robustness is a selection signal.
🔗	Claim-backed verification	A verifier matches every number in the compiled draft to a ledger measurement; the run cannot terminate while any claim is unbacked.
🗺️	Domain-aligned modification space	A seven-category 3DGS taxonomy wraps EVOLVE-BLOCK hooks that an auto-mapper places in any 3DGS repository.

Architecture

A master orchestrator (the professor) delegates to phase coordinators (PhD students) and atomic leaf agents (undergraduates):

Prompt / paper URL / GitHub repo / nothing
  │
  ├─ [1] Grounding ────────── regime classifier · literature corpus + arXiv gap-fill
  │                           repo-reproducer clones + reproduces the target (±0.2 dB)
  │                           evolve-block-mapper places the 9 EVOLVE-BLOCK hooks
  │
  ├─ [2] Grounded ideation ── bottleneck-analyst (empirical B-IDs) + limitation-brainstormer
  │                           (conceptual L-IDs) → ideator + web-idea-scout
  │                           idea-level Elo pre-selection, novelty check, tweak-class ideas banned
  │
  ├─ [3] Implementation ───── implementer: parse → plan → code → review → validate → integrate → test
  │                           debugger / freeform-mutator on failure
  │
  ├─ [4] Tournament ───────── smoke (500 it) → single-scene probe → multi-scene → full 30k
  │                           cascade-weighted Elo · falsifier · 8-island evolutionary search
  │
  ├─ [5] Integrity audit ──── ablation-runner · failure-case-analyst · provenance / novelty / claim checks
  │
  └─ [6] Manuscript ───────── figure-generator · summary/body/related-work writers · latex-assembler
                              3 parallel reviewers → revision-agent until converged → paper.pdf

Stage	Subagents
Grounding	regime-classifier · researcher · pdf-harvester · repo-reproducer · evolve-block-mapper · baseline-runner · benchmark-published
Ideation	bottleneck-analyst · limitation-brainstormer · ideator · web-idea-scout · devils-advocate
Implementation	implementer · debugger · freeform-mutator · smoke-trainer · trainer
Analysis	ablation-runner · failure-case-analyst · failure-post-mortem · visual-comparator
Writing	figure-generator · summary-writer · body-writer · related-work-writer · supplementary-writer · latex-assembler · paper-quality-checker
Review	reviewer-a · reviewer-b · revision-agent · rebuttal-writer

Each is a Markdown file with YAML frontmatter in .claude/agents/. See CLAUDE.md for roles, model tiers, and the two execution tracks (A: reproduce + improve a SOTA repo · B: local gs_template).

Results

Each SOTA row is the target reproduced on our GPUs under the reproduce-first contract, not copied from its paper.

Setting	Method	Limitation attacked / fix	PSNR ↑	SSIM ↑	LPIPS ↓
Nebula (180 fr)	3DGS	pinhole ignores OPENCV distortion; SfM yields ≈0 stars	32.58	0.914	0.278
	Nebula3DGS-v2	3DGS-MCMC sampler + saturation-driven star seeder	36.36 (+3.78)	0.945	0.201
Bubble (200 fr)	3DGS	SH cannot represent thin-film iridescence	27.40	0.857	0.337
	BubbleGS	MCMC chassis + Snell/Fresnel two-pass shell renderer	30.11 (+2.71)	0.906	0.224
Mip-NeRF 360 (7 sc)	FastGS	{0,1}-thresholded pixel count is frequency-blind	29.12	0.858	0.217
	TileFocalGS	rank-graded per-tile weights, EMA-smoothed across views	29.69 (+0.57)	0.869	0.194
Mip-NeRF 360 (9 sc, 109 s)	FastGS	opacity reset destabilises late refinement	25.65	0.751	0.311
	FreezeGS	freeze geometry, disable opacity reset in phase 2	26.08 (+0.43)	0.761	0.300
Mip-NeRF 360 (7 sc, 50 s)	splatfacto	severely under-converged at 50 s	20.53	0.593	0.505
	View-Curriculum-GS	first 15 s on the 30% most pose-similar views	23.36 (+2.83)	0.686	0.412
Mip-NeRF 360	MSAA-3DGS	multi-sample AA blurs fine contrast

The full table (11 reproduced targets, 11 wins) is in the paper. 3 SplatLab-written papers submitted to a CVPR 2026 workshop were accepted by human reviewers. In a blind expert study, SplatLab ideas scored higher on novelty than PhD-student ideas (5.71 vs 4.92, p = 0.041).

Papers written by SplatLab

Each first page below comes from a real compiled PDF produced by one autonomous run: the method was invented, trained, ablated, illustrated, written, and revised with no human editing. Click a page to read the paper.

All 16 papers, plus the OpenReview decisions, are on the project page.

Quick Start

Requirements

• Claude Code (claude on PATH) with a Claude Max plan or an API key
• tmux, python3 + pyyaml
• For real training runs: CUDA GPU(s), a nerfstudio conda env (ns-train / ns-eval), mineru (PDF → Markdown), TeXLive (pdflatex, bibtex)
• Optional: node ≥ 18 for the web dashboard; PaperBanana for figure generation (export PAPERBANANA_DIR=/path/to/PaperBanana)

Install

git clone https://github.com/seemandhar/SplatLab.git
cd SplatLab

conda create -n nerfstudio python=3.10 -y && conda activate nerfstudio
pip install nerfstudio pyyaml
pip install -e splatify/gs_template        # local nerfstudio-compatible 3DGS template (Track B)

# datasets go under data/ (paths are set in config.yaml → data_roots)
mkdir -p data && ln -s /path/to/mipnerf360 data/mipnerf360

A) Terminal: /research (recommended)

The classic way to run SplatLab, and the one used for every run in the paper. Start Claude Code in the repo root and the main session becomes the lab PI:

claude
> /research "frequency-aware adaptive densification for 3DGS" --hours 24
> /research-status      # live progress: thesis, idea pipeline, best result, budget
> /research-stop        # graceful halt; in-flight experiments finish
> /research-resume      # pick up after a stop or crash, from the ledger checkpoint

Fully detached in tmux, so it survives ssh disconnects:

bash scripts/launch.sh "my idea" 24        # starts claude + /research in a tmux session
tmux attach -t splatlab         # watch the master console

Ways to start a run:

/research "your idea here" --hours 24                                   # from a research prompt
/research "improve it" --paper-url https://arxiv.org/abs/2505.05587     # target a published paper
/research "improve it" --git-url https://github.com/<org>/<3dgs-repo>   # clone + reproduce a repo first
/research --hours 24                                                    # no idea given: falls back to config.yaml

Other flags: --scenes a,b,c, --max-budget-usd N, --model-preset optimized|all_opus|all_sonnet.

B) Web dashboard (alternative)

A zero-dependency Node dashboard on top of the same commands. It is handy for watching a run from a browser or a laptop:

bash start-ui.sh                            # → http://localhost:8420

＋ New Run → enter an idea and a time budget → Launch. Tabs: Overview (status, thesis, ideas, best result, milestones, budget) · Experiments (smoke → probe → multi-scene cascade) · Live (event feed + read-only master console) · Paper (compiled PDF, figures, reports) · Runs (inspect or resume any past run). Stop / Resume / Kill sit in the top bar.

UI control	Command	What happens
Launch Run	/research "idea" --hours N	Seeds the idea into config.yaml, starts claude in the splatlab tmux session, sends /research.
Overview	/research-status	Reads the ledger; shows status, thesis, ideas, best, milestones, budget.
Stop	/research-stop	Sets status: stopping; the master halts gracefully at its next iteration.
Resume	/research-resume	Starts a session and continues the most recent run from its ledger checkpoint.
Kill	tmux kill-session	Hard-stops the session; the ledger survives, and Resume picks it back up.

The dashboard keeps no run state of its own. Every view is derived from workspace/run_<ts>/ledger.json, events.jsonl, and tmux, so it is refresh-safe and crash-safe, and you can switch between the terminal and the browser mid-run.

Remote access (e.g. from a laptop):

HOST=0.0.0.0 GSS_TOKEN=$(openssl rand -hex 16) bash start-ui.sh
# open http://<your-box>:8420 — the page asks for the token once

See ui/README.md for the full API and design notes.

Configuration

Edit config.yaml (the dashboard's launch form writes to it too):

hours: 24                      # wall-clock cap (authoritative)
max_budget_usd: null           # null on Claude Max; a number for API billing
model_preset: all_opus         # optimized | all_opus | all_sonnet
dataset: mipnerf360
scenes: [garden, bicycle, stump, room, counter, kitchen, bonsai]
gpu_indices: [0, 1, 2, 3]
allow_repo_reproduction: true  # false → restrict to the local splatify/gs_template (Track B)

Project Structure

SplatLab/
├── README.md
├── CLAUDE.md                # architecture, team roster, execution tracks
├── config.yaml              # run configuration
├── run.sh · resume.sh · start-ui.sh
├── .claude/
│   ├── commands/            # /research, /research-status, /research-stop, /research-resume
│   ├── agents/              # 32 subagent definitions + shared protocols
│   └── skills/              # academic-voice cleanup (humanizer)
├── splatify/                # evolutionary search: Elo, cascade, falsifier, EVOLVE-BLOCKs
│   └── gs_template/         # nerfstudio-compatible 3DGS template + method plugins
├── scripts/                 # launch, ledger helpers, GPU reservation, watchdog, metric extraction
├── ui/                      # optional zero-dependency Node dashboard (server.js + public/)
├── latex/ · refs/           # paper template + BibTeX corpus
├── assets/                  # figures used in this README
└── discoveries.txt …        # reference corpora read by the ideator and writers

Output

workspace/run_<timestamp>/
├── ledger.json          # durable state: ideas, experiments, Elo, claims → measurements
├── events.jsonl         # append-only event log
├── sota/                # cloned + reproduced target repo, one worktree per idea
├── figures/             # teaser, pipeline, qualitative renders, tables
├── paper/main.pdf       # the compiled manuscript
└── RESEARCH_LOG.md      # written when no result clears the bar

Responsible Use

A tool that compresses months into days can flood a peer-review system built for human pace. We recommend labelling derivative submissions as machine-generated and shipping their experiment ledger, and we discourage volume submissions. SplatLab is a research instrument, not a paper-mill backend.

Runs consume real GPU time (and API spend when not on Claude Max). The master may clone repos and run pip/conda to reproduce baselines; set allow_repo_reproduction: false to keep it on the local gs_template. The dashboard binds 127.0.0.1 by default. Never bind 0.0.0.0 without setting GSS_TOKEN.

Related

• NERFIFY (CVPR 2026 Highlight): a multi-agent framework that turns NeRF papers into trainable Nerfstudio code.
• SPLATIFY (arXiv 2610.09116): reproduce, discover, innovate. A multi-agent framework that turns Gaussian Splatting papers and one-line ideas into trainable gsplat / Nerfstudio code. SplatLab builds on its splatify/ evolutionary search and gs_template.

Citation

If you find this work useful, please cite:

@misc{jain2026splatlabautonomousscientificdiscovery,
      title={SplatLab: Autonomous Scientific Discovery for 3DGS Research}, 
      author={Seemandhar Jain and Manmohan Chandraker},
      year={2026},
      eprint={2610.30468},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2610.30468}, 
}