Risk Masking
A mechanism that dynamically computes and protects sensitive parameter subspaces of a model, balancing editing against stability.
Timeline (last 90 days)
The ROSE framework contains two key designs: orthogonal subspace navigation and a risk-guided mechanism, the latter dynamically computing risk masks to protect sensitive parameter subspaces.
Relying solely on orthogonal constraints may be too rigid, because certain parameter subspaces are extremely sensitive to the model's overall stability.
All Facts (2)
The ROSE framework contains two key designs: orthogonal subspace navigation and a risk-guided mechanism, the latter dynamically computing risk masks to protect sensitive parameter subspaces.
50%UnverifiedRelying solely on orthogonal constraints may be too rigid, because certain parameter subspaces are extremely sensitive to the model's overall stability.
50%