Risk Evaluation · Benchmark

PaperBench: Evaluating AI's Ability to Replicate AI Research

NUWA research note · independent review pending

First published: 2025-04-02Source checked: 2026-10-25

Original source ↗Cite this record ↓Autonomy →

NUWA Research Note

Research question Design and assumptions Source findings NUWA interpretation Limits and contrary evidence Research and decision relevance Institutions and roles Original resources

Model-assisted reading · Selected sections of the original full text or official article; a complete section-by-section audit remains pending.

At a glance

NUWA interpretation: rubric-weighted partial replication is distinct from the fraction of papers fully replicated.

Research question

Can agents reproduce machine-learning papers from scratch rather than merely write plausible implementations?

Design and assumptions

Twenty ICML 2024 papers are decomposed into 8,316 graded requirements, with author-developed rubrics and an LLM judge tested on JudgeEval.

Source findings

Claude 3.5 Sonnet (New) obtained a mean replication score of 21.0%; o1 obtained 13.2% in the main experiment.

Human PhD baselines were collected on a subset, not the complete paper population.

NUWA interpretation

NUWA interpretation: rubric-weighted partial replication is distinct from the fraction of papers fully replicated.

Limits and contrary evidence

Paper selection, automated grading and subset-only human baselines constrain broad research-automation claims.

Research and decision relevance

Model recommendation: distinguish code-development requirements, executed experiments and full replication before reuse.

Precise source locations

PaperBench: Evaluating AI's Ability to Replicate AI Research · AbstractSelected original resource https://arxiv.org/html/2504.01848v3; sha256:405cc40108adb6b7c7453401ad65c6ef465e2cbc86e896ee3338350fb4f26ac7; archived source capture; exact capture time not independently established
PaperBench: Evaluating AI's Ability to Replicate AI Research · §3 DatasetSelected original resource https://arxiv.org/html/2504.01848v3; sha256:405cc40108adb6b7c7453401ad65c6ef465e2cbc86e896ee3338350fb4f26ac7; archived source capture; exact capture time not independently established
PaperBench: Evaluating AI's Ability to Replicate AI Research · §4 LLM JudgeSelected original resource https://arxiv.org/html/2504.01848v3; sha256:405cc40108adb6b7c7453401ad65c6ef465e2cbc86e896ee3338350fb4f26ac7; archived source capture; exact capture time not independently established
PaperBench: Evaluating AI's Ability to Replicate AI Research · §5.2 Main ExperimentSelected original resource https://arxiv.org/html/2504.01848v3; sha256:405cc40108adb6b7c7453401ad65c6ef465e2cbc86e896ee3338350fb4f26ac7; archived source capture; exact capture time not independently established
PaperBench: Evaluating AI's Ability to Replicate AI Research · §5.4 Human Baseline PerformanceSelected original resource https://arxiv.org/html/2504.01848v3; sha256:405cc40108adb6b7c7453401ad65c6ef465e2cbc86e896ee3338350fb4f26ac7; archived source capture; exact capture time not independently established

Source checking, extraction, editorial approval and replication are recorded separately. Methodology →

Institutions and roles

Research by OpenAI

Authors (13)

Giulio Starace · Oliver Jaffe · Dane Sherburn · James Aung · Jun Shern Chan · Leon Maksin · Rachel Dias · Evan Mays · Benjamin Kinsella · Wyatt Thompson · Johannes Heidecke · Amelia Glaese · Tejal Patwardhan

Evaluation protocols

PaperBench →

Cite and reuse

Cite the original publication for its findings. Cite this NUWA record separately for its explanation and retained snapshot.

BibTeX ↓RIS ↓Record + source metadata · JSON ↓Reading coverage · JSON Suggest a correction

Original publication · BibTeX

@misc{paperbench,
  title = {PaperBench: Evaluating AI's Ability to Replicate AI Research},
  url = {https://arxiv.org/abs/2504.01848v3},
  author = {Giulio Starace and Oliver Jaffe and Dane Sherburn and James Aung and Jun Shern Chan and Leon Maksin and Rachel Dias and Evan Mays and Benjamin Kinsella and Wyatt Thompson and Johannes Heidecke and Amelia Glaese and Tejal Patwardhan},
  year = {2025},
  date = {2025-04-02},
  version = {arXiv:2504.01848v3},
  note = {arXiv:2504.01848v3}
}

NUWA record · BibTeX

@misc{paperbench-nuwa,
  title = {PaperBench: Evaluating AI's Ability to Replicate AI Research — NUWA research record},
  url = {https://whitzard.tech/en/nuwa/whitzard-index/work/paperbench/},
  version = {2026-10-29.1},
  organization = {NUWA Frontier AI Safety Lab},
  howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--paperbench.json},
  note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--paperbench.json.}
}

Evidence snapshot 2026-10-29.1 · Index updates · RSS

References

[1] Starace, Giulio, Jaffe, Oliver, et al.. PaperBench: Evaluating AI's Ability to Replicate AI Research. 2025-04-07. v3 ; Author affiliations on the original paper title page; Abstract; §3 Dataset