MP-IB
MP-IB achieved the best Bridge2AI agitation prediction among the evaluated methods. MP-IB's novelty is heterogeneous arithmetic precision across semantic heads rather than ordinary layer compression for one task. MP-IB uses numerical preci…
1 sources - 5 claims
MP-IB achieved the best Bridge2AI agitation prediction among the evaluated methods. MP-IB's novelty is heterogeneous arithmetic precision across semantic heads rather than ordinary layer compression for one task. MP-IB uses numerical precision as an information bottleneck to separate stable speaker traits from volatile agitation-related state information. MP-IB uses an FP16 64-dimensional trait head and an INT4 32-dimensional state head with an 8x capacity asymmetry. MP-IB's advantage is interpreted as coming from T-MAE pretraining, a low-capacity INT4 state bottleneck, and asymmetric trait-state precision.