DualMLC Framework Combines Decoder-Only and Bidirectional Models for Multi-Label Classification
September 14, 2026
DualMLC uses a dual-branch architecture that processes documents through both an autoregressive decoder-only model and a bidirectional encoder. The framework employs late logit fusion to combine independent relevance scores, improving classification accuracy across large-scale label spaces.
HOW THIS AFFECTS YOU
●
builderThis architecture may improve classification performance in systems handling tens of thousands of labels.
●
researcherThe late logit fusion approach explores the complementarity of heterogeneous model architectures.