TMLR Editors Quizzed Authors on Their Own Papers — Most Couldn't Explain Them

An editor-in-chief of Transactions on Machine Learning Research (TMLR) ran an informal experiment: rather than silently desk-rejecting a batch of ten weak submissions, he invited the authors to a live meeting to talk through their own papers. The attendees spanned the academic spectrum — undergraduates, master’s and PhD students, faculty, and independent researchers — and most submissions were solo-authored. The results were damning. Authors of several papers could not answer even basic questions about their work, and others could gesture at the high-level idea but fell apart the moment the questions turned to technical specifics. Only a small minority demonstrated genuine command of what they had supposedly written.

The backdrop is the flood of likely LLM-generated submissions now hitting ML venues. TMLR’s desk-rejection rate has climbed from roughly 6% in 2023 to around 53%, and the interviews put a human face on that statistic: papers that look polished on the page but whose ‘authors’ show no real engagement with the content. A telling detail is that some authors who froze during the live conversation later followed up with detailed written explanations — exactly the pattern you’d expect if a language model, not the person, was doing the reasoning.

The episode reframes AI-driven academic fraud as an authorship-verification problem, not just a plagiarism or quality problem. It fuels an ongoing debate about whether journals should adopt lightweight oral defenses to confirm authors actually understand their submissions — a check that catches ghost-written and machine-generated work but doesn’t obviously scale to the volume peer review now faces.