Evidence record

PaperBench

Original paper v1 · Gemini 2.0 Flash / BasicAgent · 2025-04-02

Correction or update linkedCleared a metadata field that repeated the standard-error amplitude as if it were a confidence level; printed values and reported standard errors are unchanged.

Reported result · numeric

3.2 ± 0.2%

Standard error

Mean replication score · percent

benchmark author reported · extraction review: agent checked

Source and extraction

Published 2025-04-02

PaperBench: Evaluating AI’s Ability to Replicate AI Research (v1)Section 5.2, Table 4Original source ↗

Evaluation setup

task count	20
run count	3
aggregate method	Mean replication score; standard error across runs
wall clock budget	12 hours
internet access	true
scaffold	BasicAgent
task exclusions	Author code repositories blacklisted

Not reported: split, task snapshot, attempts per task, selection rule, token budget, hardware, training budget, inference budget, monetary cost, tool access, filtering, evaluator version, human intervention, contamination concerns, comparability caveats.

Comparability

direct comparison

Point differences do not establish significance.

Limitations

Rubric credit is not the fraction of papers fully replicated.
Scaffold changes can reverse model ordering.