ImpossibleRubrics: Stress-Testing Generated Rubrics as Reward Signals
Tailored AI-written rubrics can make impossible prompts easier to game, not harder.
The paper introduces ImpossibleRubrics, a 169-task benchmark where the honest answer is to reject an unsupported conclusion. Eleven rubric generators were exploitable 8–26% of the time on the main cut, and the strongest measured generator still failed 36% of the time on a selected stress cut. A certificate-faithful rubric had 0% exploitation, so the authors frame the gap as rubric quality rather than unavoidable task difficulty. The sharp result: seven generators did worse than a fixed generic rubric, suggesting task-specific criteria can point attackers toward the lie to fabricate. HF Daily Papers' note
The paper introduces ImpossibleRubrics, a 169-task benchmark where the honest answer is to reject an unsupported conclusion. Eleven rubric generators were exploitable 8–26% of the time on the main cut, and the strongest measured generator still failed 36% of the time on a selected stress cut. A certificate-faithful rubric had 0% exploitation, so the authors frame the gap as rubric quality rather than unavoidable task difficulty. The sharp result: seven generators did worse than a fixed generic rubric, suggesting task-specific criteria can point attackers toward the lie to fabricate. HF Daily Papers' note
score 5