Agreement-Before-Diversity: A New Decision Rule for Heterogeneous Language-Model Ensembles

A recent paper published on arXiv (2608.04618) presents a novel decision-making framework called Agreement-Before-Diversity (ABD) for managing diverse language-model ensembles. This approach separates the authority for replacement from candidate headroom, allowing for explicit and auditable oversight. ABD operates as a static, label-free guideline: an anchor response is kept if it is validated by two other reliable samples under a predetermined equivalence relation; if not, it is substituted with a varied synthesis. The authors establish two precise identities: the accuracy gap in relation to unconditional synthesis is influenced by agreement coverage and the anchor's advantage within the protected subset, while the gap concerning the absence of synthesis contrasts authorized recovery with authorized destruction. This methodology tackles the absence of a clear standard for when a new generated response should replace an existing one in heterogeneous ensembles. The full paper can be accessed at https://arxiv.org/abs/2608.04618.