ReTeX Recovers Task-Expert Performance from Merged Checkpoints Without Storing Separate Experts

ReTeX models parameter interference from multi-task model merging as affine transformations approximated as additive offsets, then trains a predictor to undo them at inference from a single merged checkpoint. A router-free task identifier based on SVD subspace signatures handles unknown task identity, avoiding the storage overhead of dynamic merging approaches.