MultiRisk: Multiple Risk Control via Iterative Score Thresholding

The study proposes the MultiRisk test-time filtering framework, which uses iterative score thresholding to simultaneously control multiple risks of generative AI. The method formalizes multi-risk constraints with user priorities, and proposes two dynamic programming algorithms, MULTIRISK-BASE and MULTIRISK; the latter leverages data exchangeability to guarantee that each risk is simultaneously controlled. On the three-constraint large language model alignment task on the PKU-SafeRLHF dataset, each individual risk is controlled near its target level.