Swiss-Knife: A reconfigurable externalized multi-objective alignment framework at decoding time

The researchers propose Swiss-Knife, an externalized multi-objective alignment framework at decoding time that turns alignment specifications into first-class runtime objects, composed of hot-swappable scoring blades, a batch normalizer, a pairwise aggregation operator, and a selection rule. The authors argue that existing decoding-time alignment, which scores candidate continuations of a frozen language model with external rewards and selects the maximum, is merely a degenerate point in a larger design space. The paper characterizes admissible aggregation operators with six axioms and proves a representation theorem, in which pairwise aggregation remains Lipschitz stable under adversarial reward contamination while argmax does not, and Candidate-Batch Normalization (CBN) makes the weight simplex invariant to reward model rescaling.