SiM enables dynamic routing in multi-task model merging without additional training or task ID access. It uses SVD-based manifold approximations and projects test inputs onto precomputed task manifolds to route inputs to relevant experts, improving performance and reducing the gap to individual expert levels.