CORE-MTL: Rethinking Gradient Balancing via Causal Orthogonal Representations

Researchers introduced CORE-MTL, a multi-task learning framework that separates shared features into semantic and residual streams.

Instead of gradient reweighting or projection, it uses representation factorization and domain-specific priors for vision tasks.

The method is designed to reduce negative transfer and improve both in-distribution and out-of-distribution generalization.

The authors report a stronger OOD bound and better benchmark results than prior multi-task learning approaches.