PaperlayerView exact revision on arXiv

Source-linked AI summary

SplatLab: Autonomous Scientific Discovery for 3DGS Research

Seemandhar Jain, Manmohan Chandraker

arXiv:2610.30468v1cs.CVcs.AI

TL;DR

Reproducing 3DGS papers is costly, and prior autonomous-research systems have not demonstrated a verified end-to-end pipeline in an active computer-vision subfield. SplatLab builds on verified reproductions and links manuscript claims to logged GPU runs; its PSNR-targeted manuscripts beat their reproduced targets, while the research cycle compresses from months to days of mostly autonomous compute.

  • Problem

    Reproducing a 3DGS paper can take weeks of expert engineering, and prior autonomous-research systems lack an end-to-end demonstration in a live computer-vision subfield with verified baselines and claim-to-measurement traceability.

    Source: §1 · p. 2

  • Method

    SplatLab evolves plugins verified against published 3DGS targets, evaluates candidates with cascade-weighted Elo and in-loop falsification, and binds reported numbers to logged runs.

    Source: §1 · pp. 2–3, §3 · p. 4

  • Results

    +0.18 to +3.78 dB PSNR: SplatLab manuscripts beat their reproduced targets on Mip-NeRF 360 and LLFF, including +3.78 dB on astrophotographic nebula rendering.

    Source: §1 · p. 3

  • Takeaways & Limitations

    The research cycle that takes graduate students months compresses to days of mostly autonomous compute.

    Source: §5 · p. 10

    Building a verified-reproduction backbone for each field is costly, biasing capability toward pre-engineered subfields and well-resourced labs.

    Source: §5 · p. 10

Abstract

from arXiv · show

Reproducing a published 3DGS paper takes weeks of expert engineering before any extension can be proposed. Prior autonomous-research systems read papers, draft hypotheses, and write code, but are limited by short fixed templates, low executability, and weak traceability, and none has been demonstrated end-to-end in a live computer-vision subfield with intense activity. SplatLab is an autonomous research system for 3D Gaussian Splatting that addresses these limitations. Given a research prompt, a target-paper URL, or no input at all, it reproduces a published baseline, proposes an extension, trains and ablates across Mip-NeRF 360 and LLFF, writes the manuscript, and binds every reported number to a logged training run. Five design choices distinguish it: (i) candidates mutate plugins from a verified-reproduction backbone matched to published baselines; (ii) a cascade-weighted Elo tournament across eight islands scores candidates on measured GPU PSNR with verbal-gradient feedback; (iii) a standing falsifier subtracts Elo on failure, integrating stress-test survival into selection; (iv) every reported number is bound to a logged training run, and the run cannot terminate while any claim is unbacked; (v) a seven-category 3DGS taxonomy wraps EVOLVE-BLOCK markers in domain-aligned scaffolds transferable across 3DGS repositories. 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. Code, all manuscripts, experiment ledgers, and the cross-run paper store are available on the project page.

1 Introduction

SplatLab targets the costly, weakly verified reproduction step that limits autonomous research in active computer-vision subfields. It builds on verified 3DGS reproductions and combines measured GPU evaluation, falsification, and claim-to-run traceability; its manuscripts exceed reproduced targets on PSNR.

Motivation

  • Reproducing a published method can require engineering effort comparable to producing a new one, while researcher time and reading bandwidth do not scale with subfield publication rates.The authors are unaware of an end-to-end autonomous-research pipeline demonstrated in a computer-vision subfield where reproducing one paper takes weeks of expert engineering.Source: §1 · p. 2

Approach

  • SplatLab mutates SPLATIFY plugins already verified to train and match published 3DGS targets, giving proposed extensions a fixed reproduction baseline.SPLATIFY extends NERFIFY with a modular template and nine composable extension hooks.Source: §1 · pp. 2–3
  • A cascade-weighted Elo tournament ranks candidates using measured GPU PSNR, while an in-loop falsifier deducts Elo when candidates fail stress tests.Verbal-gradient feedback from tournament outcomes is fed back to the ideator.Source: §1 · pp. 2–3
  • Every reported number is bound to a logged experiment, and the run cannot terminate while a claim lacks a backing measurement.The system also uses seven 3DGS novelty categories with EVOLVE-BLOCK scaffolds transferable across repositories.Source: §1 · p. 3

Results

  • +0.18 to +3.78 dB PSNR: SplatLab manuscripts beat their reproduced targets on Mip-NeRF 360 and LLFF, including +3.78 dB on astrophotographic nebula rendering.The nebula domain had not been addressed by prior 3DGS work.Source: §1 · p. 3

2 Related Work

Prior autonomous-research work includes systems that produce manuscripts, systems that run discovery loops without papers, and evolutionary program-search methods. SplatLab combines verified-reproduction substrates with grounded evaluation and reviewer-triggered experimentation, distinctions the authors draw against these lines of work.

End-to-end automated research pipelines

  • Manuscript-producing systems use templates, agentic research loops, or review mechanisms, but the surveyed systems build on agent-written reimplementations not verified against their targets.SplatLab requires candidate plugins to compile, train, and reach the target’s headline; it also allows reviewer-triggered re-experimentation.Source: §2 · p. 3

Autonomous discovery without manuscript output

  • Discovery systems without manuscript output include biomedical Elo tournaments and experimentally validated laboratory work, while ideation studies assess research ideas separately from a full paper-producing loop.The cited study of more than 100 NLP researchers found LLM ideas judged more novel but slightly less feasible than expert ideas [36].Source: §2 · p. 3

Evolutionary program search and multi-agent infrastructure

  • SplatLab adopts ReEvo’s verbal gradients but routes them to the ideator and uses cascade-weighted, grounding-conditional Elo rather than per-individual fitness.The authors contrast its compiled training runs against a verified substrate with Elo ranking of free-text hypotheses and uniformly updated ratings.Source: §2 · p. 3

3 Method

SplatLab performs iterative research over a verified 3DGS reproduction substrate, producing manuscripts and code with reported numbers linked to training runs. Its stages ground research in measured bottlenecks and literature, evolve plugin mutations through GPU evaluation and falsification, and audit and revise the resulting manuscript.

3 Method

  • SplatLab maps a prompt, target-paper URL, or no input to manuscript source and PDF, a ledger linking reported numbers to runs, and the code that produced them.The fixed context includes the 3DGS literature corpus, SPLATIFY backbone, nine EVOLVE-BLOCK hooks, and target conference template.Source: §3 · p. 4

3.1 SplatLab

  • Grounding reads prior work and trains the verified plugin on representative scenes, turning literature limitations and measured bottlenecks into the thesis spine for later stages.The bottleneck analysis profiles per-tile loss, PSNR trajectory, memory, and gradient norms.Source: §3.1 · p. 4
  • Grounded ideation ranks candidate mechanisms tied to literature limitations and measured bottlenecks, filtering for precedent, novelty, category saturation, and orthogonality before plugin implementation.Candidates receive an initial Elo rating and are compared on novelty, bottleneck specificity, implementability, and predicted-delta plausibility.Source: §3.1 · p. 5
  • Implementation turns surviving ideas into structured diffs within EVOLVE-BLOCK hooks on verified SPLATIFY or NERFIFY plugins, then checks compilation and a 100-iteration smoke test.Failures return to the Implementer with build logs as feedback, subject to a repair budget.Source: §3.1 · p. 5
  • The tournament scores candidates on real PSNR with Elo conditioned on the benchmark-and-loss grounding pair, and weights updates by evaluation stage so smoke wins cannot outweigh multi-scene losses.The stages progress from smoke to probe, multi-scene, and full benchmark; only the champion enters the full benchmark.Source: §3.1 · p. 5
  • A standing falsifier attacks survivors on edge scenes, sparse-view reduction, and published-range hyperparameter perturbations; each failed attack subtracts 30 Elo before manuscript writing.Survivors are spread across eight islands, with promotion requiring Pareto improvement over the benchmark-and-loss pair.Source: §3.1 · p. 6
  • The integrity audit checks provenance, rationale-to-diff faithfulness, novelty, compute matching, cross-codebase portability, and whether every claim is supported by the experiment ledger.Load-bearing claim mismatches terminate the run.Source: §3.1 · p. 6
  • Manuscript synthesis combines writers, figures, ablations, and reviewer agents, with reviewer concerns able to trigger new experiments and re-entry into implementation or refinement.The loop ends when no reviewer raises a new-experiment concern.Source: §3.1 · p. 6

4 Experiments

SplatLab is evaluated through end-to-end runs, benchmark performance, manuscript quality, peer review, and component ablations. Its discovered methods improve PSNR over paired baselines, while evaluations also examine robustness, claim traceability, and how generated papers compare with published work.

Run outcomes

  • Across 16 runs, SplatLab trained 198 experiment branches, producing five verified positives, five marginal results, five non-PSNR contributions, and one recorded failure.Runs used 2–36 GPU-hours each and estimated LLM API costs of $150–250.Source: §4 · p. 7

Generated Methods vs. SOTA

  • +0.18 to +3.78 dB PSNR gains over paired baselines span standard 3DGS benchmarks and novel scenes, with the largest gain on Nebula.Table 1 presents a representative subset; the full table is in the Appendix.Source: §4 · pp. 7–8, Table 1 · p. 7

Multi-seed Robustness

  • +0.23 ± 0.09 dB, GNS-AA reaches 26.18 ± 0.05 dB versus GNS at 25.95 ± 0.07 dB across five seeds.The original reported gain was +0.18 dB; improvements within the estimated noise floor are classified as marginal rather than verified.Source: §4 · p. 8

SplatLab vs. Prior Discovery Systems

  • Only SplatLab satisfies all five capability axes compared across the eight closest autonomous-research and discovery systems.The comparison includes a verified-reproduction backbone, grounding-conditional idea re-ranking, an in-loop falsifier, claim-backed verification, and reviewers that can trigger training runs.Source: §4 · p. 8, Table 2 · p. 8

Idea and Manuscript Quality

  • SplatLab outperforms prior systems by large margins on ideation and scores close to human ground truth at manuscript level.Idea evaluations compare 20 candidates with student and accepted-paper ideas; manuscript evaluations use five human-verified prompts and graphics-area reviewer verification.Source: §4 · p. 8, Table 3 · p. 8, Table 4 · p. 9

LLM and Human Peer Review

  • Human reviewers score published papers higher in Soundness and Presentation, but SplatLab higher in Novelty; the multi-agent reviewer slightly favors SplatLab on every axis.Published papers score 4.1 overall versus 3.5 for SplatLab; the authors report that machine-pace sweeps surface mechanisms human-paced research has not yet reached.Source: §4 · pp. 8–9, Table 5 · p. 9

Real Workshop Acceptance

  • Three SplatLab-generated manuscripts were accepted as posters at a CVPR 2026 workshop.The manuscripts were AttDis & Friends, NeedleSwap, and AuditSplat.Source: §4 · p. 9

Component Validation and Ablations

  • Without claim-backed verification, every ablated manuscript included a number untraceable to a disk log; the full pipeline produced zero such cases.Ablations reuse the canonical-run winner and the same 32-idea / 16-paper pool.Source: §4 · p. 9, Table 6 · p. 9

4.3. Feature ablations

The taxonomy ablation suggests domain grounding improves search efficiency but is not necessary for execution; the figure caption also reports three generated papers accepted as CVPR 2026 workshop posters.

4.3. Feature ablations

  • AttDis & Friends, NeedleSwap, and AuditSplat were accepted as posters at a CVPR 2026 workshop.Figure 4 shows pages 1–8 of each paper from left to right.Source: §4.4 · p. 10
  • +3.78 dB and +3.45 dB PSNR were achieved by taxonomy-guided and taxonomy-free runs, respectively, on the nebula target.The single-target comparison suggests taxonomy grounding improves search efficiency rather than being strictly necessary for execution.Source: §4.4 · p. 10
  • Removing the taxonomy increased token consumption from 47.5M to 85.5M (80%) on the nebula target.Stages 1 and 3–6 are designed to operate across domains, while Stage 2 provides domain-specific ideation grounding.Source: §4.4 · p. 10

5 Conclusion

SplatLab demonstrates autonomous research in a live computer-vision subfield, with verified reproductions and claim-to-run binding; it also raises risks around review load, unequal access, and research opportunities.

5 Conclusion

  • SplatLab produces submission-ready 3DGS manuscripts that beat their reproduced targets on PSNR, while compressing a months-long research cycle to days of mostly autonomous compute.The conclusion says a single research cluster can sweep design-space slices that hand-execution would take years to cover.Source: §5 · p. 10
  • SplatLab’s openness and speed may flood human-paced peer review, while backbone-building costs favor well-resourced labs and pre-engineered subfields.A bug in the integrity audit could mask hallucination rather than catch it.Source: §5 · p. 10
  • The authors recommend labeling derivative submissions and shipping experiment ledgers, and propose using Mode A logs for education.They describe SplatLab as a research instrument, not a paper-mill backend.Source: §5 · p. 10

Related papers

Loading 2610.30468v1…