SMIC Improves NLI via Multiscale Input-Reference Conditioning
November 13, 2026
SMIC uses a single input-dependent multiscale reference shared across Transformer layers to improve Natural Language Inference. The token-context Transformer (TCT-SMIC) achieved the highest mean accuracy across all three ANLI rounds.
HOW THIS AFFECTS YOU
●
researcherYou can use this conditioning method to make models less sensitive to small lexical changes in text.