Trajectory-Invariant Latent Refinement (TILR)

A training-free intervention that extracts a low-rank invariant subspace via SVD of contrastive differences between a later- and earlier-checkpoint's latent reasoning trajectories, then at inference projects each step's refinement update onto only that subspace (gated by a norm-based reliability check) to stabilize latent chain-of-thought.