RebuttalAgent: AI learns to "read the minds" of peer reviewers (and rescue your papers)

Getting published in top journals and at ICLR-level conferences has long become an elaborate psychological duel, one researchers themselves call "dancing in shackles." You spend years developing an architecture, weeks training models and months polishing the text, only to face the verdict of an anonymous Reviewer #2 who skimmed your paper. The rebuttal stage is the moment of truth: either you convince your opponent that you are right, or your work goes into the archive until next year. The problem is that reviewers rarely say outright what bothers them, leaving authors to guess between the lines.

Researchers at the Hong Kong University of Science and Technology decided it was time to change the rules of this game and introduced RebuttalAgent. It is not yet another ChatGPT add-on for fixing grammar, but a sophisticated system built on the principles of "theory of mind." In psychology, the term describes the ability to understand that other people have their own beliefs, intentions and knowledge, which may differ from yours. RebuttalAgent tries to reconstruct the reviewer's mental state: it analyzes the text of the review to determine whether the criticism stems from genuine misunderstanding, a lack of data or fundamental disagreement with the methodology.

The system works through multi-step analysis. First, the agent deconstructs each reviewer comment, identifying hidden cognitive biases and expectations. Then it models possible dialogue trajectories: what happens if we respond aggressively, and what if we admit a mistake. It resembles a chess engine calculating how a game could unfold, except that the pieces on the board are arguments and scientific facts. RebuttalAgent picks wording that does not merely answer the question but "closes" the critic's psychological gestalt, making their position more open to a higher score.

Why does this matter now? As submissions to AI conferences grow exponentially, review quality inevitably drops. Reviewers are overloaded, they make mistakes and often miss the point. In that environment, the winner is not the smartest author but the most persuasive one. RebuttalAgent effectively democratizes access to the "art of persuasion," which is especially important for scientists whose first language is not English.

On the other hand, we are entering a danger zone where scientific discourse turns into a battle of algorithms. If AI writes the paper, AI (often) reviews it, and now AI also defends it, where in that chain is there room for a genuine increase in human knowledge?

The industry has already seen attempts to automate paper writing, but the Hong Kong team's work goes deeper. They have taken aim at the holy of holies: academic reputation and the expert evaluation process. If RebuttalAgent proves its effectiveness over the course of ICLR 2026, we could face a crisis of trust in the institution of peer review itself. After all, if any critic can be "talked around" by a well-chosen script, such criticism is worth next to nothing.

Bottom line: the scientific community has gained a legal cheat code for getting past the gatekeeping of top conferences. Will it become a tool of fairness, or finally bury the objectivity of peer review?