Learning to Steer, Steering to See: Unveiling the Geometry of RLVR in Large Language Models via Trainable Vectors
The Alpha-Stabler framework is proposed, a plug-and-play framework with a Predictor that monitors principal-subspace intrusion for early collapse warnings, and a Controller that removes the principal-subspace component of activation gradients during backpropagation while preserving the orthogonal complement.
Yu-Chen Cai, Ding Cao, Qi-Xiang Yin et al.
· 1 citation