TADDLE: A Tool-Augmented Agent for Detecting Deficient LLM-Generated Peer Reviews

Researchers introduced TADDLE, a tool-augmented system for detecting specific deficiencies in LLM-generated peer reviews.

They also released the first expert-annotated benchmark: 1,800 reviews from 50 ICLR 2025 papers, labeled across six defect types plus a non-deficient class.

Unlike simple authorship detection, the approach focuses on concrete review quality problems, making it more useful for integrity checks and evaluation.

The work, led by Hanqi Duan and shared on arXiv, supports stronger assessment of AI-written peer reviews.