Project APE builds its verifier benchmark from 100 AI-written papers with injected errors

soumitrashukla9 · x · 2026-07-22

This post explains a core benchmark-design issue for Project APE: what counts as ground truth?

• The team starts from 100 fully AI-written papers.
• They then inject errors intentionally according to the CRED taxonomy.
• The benchmark asks whether a perfect verifier can detect all of those injected defects.

It is essentially a methodology post about how to evaluate verification systems under controlled, known errors.

Original post →

More from Research

• Open-source runtime lets each repo define its own AI code reviewer — ibabufrik · 2026-07-22
• OpenAI and Apollo Research introduce Contrastive SDF to measure reward-seeking — OpenAI · 2026-07-22
• NVIDIA says to tune the harness before tuning the model with LangChain — NVIDIAAI · 2026-07-22
• The Thimble and the Waterfall: AI's Data Bottleneck and Feedback Loops — dyamins · 2026-07-22
• NVIDIA shows 22 SIGGRAPH papers and Omniverse tools for robot simulation — facontidavide · 2026-07-22
• Building a Knowledge Graph Without a Graph DB: 1000x Cheaper Than GraphRAG — TheRedfather · 2026-07-22

Research channel →