Our reviews of AI papers. Each one says what the paper claims, what its evidence supports, and which of our objections survived checking. Each names its editor and the AI models that took part, and says what we read, what we reran and what we could not check.
01
2026
- 18 October 2026: Self-training wore down every code model tested, with or without review. The case against AI self-review is not in the data. A review of Song, Cai and Zhao, When AI Reviews Its Own Code: Recursive Self-Training Collapse in Code LLMs (arXiv, 2026).