[Paper] To Err Is Human: Systematic Quantification of Errors in Published AI Papers via LLM Analysis

Source: arXiv - 2512.05925v1

Overview

A new study leverages a cutting‑edge large language model (GPT‑5) to automatically scan AI conference and journal papers for objective mistakes—wrong formulas, mis‑drawn figures, incorrect tables, etc. By quantifying these errors across several top venues, the authors reveal that even high‑impact publications contain a growing number of verifiable bugs, and they demonstrate that an LLM can not only spot them but also suggest fixes.

Key Contributions

• Paper Correctness Checker: A GPT‑5‑based tool that parses PDFs, extracts mathematical and tabular content, and flags inconsistencies with a ground‑truth verifier.
• Large‑scale error audit: Analyzed papers from NeurIPS (2021‑2025), ICLR (2018‑2025), and TMLR (2022‑2025), uncovering an average of 5–6 objective mistakes per paper.
• Human validation: Expert reviewers confirmed 263 out of 316 flagged items (83.2 % precision).
• Automated repair: The system generated correct replacements for ≈ 76 % of verified mistakes.
• Trend insight: The mean mistake count rose by ~55 % from NeurIPS 2021 to NeurIPS 2025, suggesting quality‑control pressures are mounting.

Methodology

1. Paper ingestion – PDFs are converted to a structured representation (text, LaTeX snippets, tables, figures).
2. LLM reasoning – GPT‑5 is prompted with domain‑specific checks (e.g., “Does the derivative of f(x) match the expression in Eq. 3?”).
3. Ground‑truth verification – For each claim, a lightweight symbolic engine or statistical test validates the LLM’s suspicion (e.g., recomputing a numeric table).
4. Human audit – A panel of AI researchers reviews a random subset of flagged items to estimate precision.
5. Fix generation – When a mistake is confirmed, the same LLM is asked to produce a corrected version, which is then cross‑checked automatically.

The pipeline is deliberately limited to objective, verifiable errors; subjective judgments about novelty or writing style are excluded.

Results & Findings

• Precision: 83.2 % (263/316) of flagged items were true errors.
• Error severity: Most were minor (typos in equations, mismatched table entries), but a handful could alter result interpretation.
• Repair success: The LLM supplied a correct fix for 75.8 % of verified mistakes, often with a concise LaTeX replacement.

These numbers suggest that even elite venues are not immune to slip‑ups, and that the volume of published work may be outpacing traditional peer‑review safeguards.

Practical Implications

• Developer tooling – Integrating a similar “correctness checker” into manuscript‑authoring platforms (e.g., Overleaf plugins) could catch errors before submission.
• Reproducibility pipelines – Automated verification of equations and tables can be added to CI/CD workflows for research code, reducing downstream debugging.
• Peer‑review augmentation – Journals and conferences could deploy LLM‑based assistants to flag obvious objective mistakes, letting reviewers focus on novelty and methodology.
• Knowledge‑base hygiene – Curators of open‑source model cards, benchmark leaderboards, and literature surveys can run the checker to prune propagated errors.

In short, the study demonstrates a practical, scalable safety net that can improve the reliability of AI research without replacing human expertise.

Limitations & Future Work

• Scope of errors: The system only handles objectively verifiable issues; nuanced methodological flaws remain out of reach.
• Domain dependence: Accuracy relies on the LLM’s familiarity with the specific subfield’s notation and conventions.
• False positives/negatives: Although precision is high, recall was not measured; some mistakes may still slip through.
• Scalability of human validation: Scaling expert review beyond the sampled 316 items would be costly.

Future directions include expanding the checker to semantic consistency checks (e.g., aligning loss curves with described algorithms), integrating with version‑controlled repositories for continuous validation, and exploring multimodal verification for figures and diagrams.

Bottom line: By turning a state‑of‑the‑art LLM into a systematic proof‑reader, the authors provide a concrete path toward cleaner, more reproducible AI literature—an advance that developers, reviewers, and research managers can start to leverage today.

Authors

• Federico Bianchi
• Yongchan Kwon
• Zachary Izzo
• Linjun Zhang
• James Zou

Paper Information

• arXiv ID: 2512.05925v1
• Categories: cs.AI, cs.CL
• Published: December 5, 2025
• PDF: Download PDF