Blog
-
UP: Unbounded Positive Asymmetric Optimization for Breaking the Exploration-Stability Dilemma
IS-based RL for LLM reasoning faces an exploration-stability dilemma: pure importance sampling explodes, and the clipping used to tame it ties a token's update budget to the old policy. By formalizing the Probability Capacity and anchoring the policy to itself with a stop-gradient, Unbounded Positive...
-
Rethinking Muon Beyond Pretraining: Spectral Failures and High Pass Remedies for VLA and RLVR
Muon orthogonalizes the momentum matrix and pushes every singular value to one. This works beautifully for LLM pretraining, which is essentially next token classification on text via supervised learning. But what happens when we move along three orthogonal axes: a different modality, a different loss,...