BRACE: Bipolar Reference-Aware Calibration and Estimation for Incomplete Multimodal Learning

Real-world multimodal data are often incomplete. Existing incomplete multimodal methods mainly address this problem as evidence attenuation, by reconstructing missing views, regularising shared representations, or retrieving auxiliary context.However, partial observations can still leave a different uncertainty unresolved:even when they provide a coarse task-state anchor, they may not determine whether the latent state should be revised toward a higher or lower label. We formalize this overlooked failure mode as residual-direction ambiguity and propose BRACE( Bipolar Reference-A ware Calibration and Estimation), a framework that first builds a quality-aware unified task state from available modalities, then organizes historical state-label pairs in a global memory bank, and finally infers a gated latent correction from the ordered contrast between higher-label and lower-label reference neighborhoods, together with a warmup-then-calibration training strategy and direction-consistency regularization. Across four benchmarks spanning multimodal sentiment analysis and multimodal misinformation detection under fixed and random missingness, BRACE consistently outperforms strong baselines, improving Acc-2 by up to 7.4 points over HME under fixed missingness and remaining robust when key modalities are absent; code is available at https://anonymous.4open.science/r/Brace-94BF.