The work tracks parameter updates during narrow-domain fine-tuning rather than examining models only after misalignment appears. Directional Hessian curvature concentrates on semantic pivot tokens, while gradient-subspace analysis indicates that harmful and safe behavior separate mainly because overlap with safe gradients declines.
A mitigation projects the empirically measured harmful gradient subspace out of subsequent parameter updates. It reduces free-generation by up to 80.0% on Qwen2.5-14B-IT. In three other model families where behavioral failures were already near zero, teacher-forced tests still found the same latent harmful subspace, suggesting that ordinary generation tests can overstate safety.