Gaowen Liu

Ground-truth access in the task-generating proposer can paradoxically *accelerate* self-play collapse, suggesting that ungrounded proposers might be more stable partners for self-consistency solvers.

Muon's "one-size-fits-all" spectral whitening can cripple VLA and RL, but a high-pass spectral filter (Pion) can restore performance by suppressing gradient noise and preserving pre-trained head specialization.