Post-Grokking Collapse at the Representation-Readout Interface in Muon-Trained Transformers

Under the standard split, Muon gets hidden matrices and AdamW embeddings/output head. Muon groks modular addition faster, but its solutions do not hold. All nine configurations on $(a+b) \bmod 113$ grok and later lose generalization. Across five seeds the selected AdamW reference falls below threshold on four, reaching 27.59%. Instability persists across two moduli, two widths, two training fractions, subtraction, and depth. The failure arises at the representation-readout interface, identified only jointly up to an invertible map unselected by the loss. After solving the training set, the gradient falls to order $10^{-6}$ and the optimizers respond differently: step-size elasticity is -0.03 for Muon versus +1.5 for AdamW, and the Muon group moves 8.0 times faster per parameter. From bit-identical states, freezing either group prevents failure. Freezing embeddings/readout removes it in five runs over 451,400 post-grokking steps and five paired seeds: unfrozen arms record 137-321 sub-threshold evaluations, frozen arms none. Removing Muon's normalization and orthogonalization is no substitute: it collapses representation from 326 effective conjugate pairs to 4, shows no recurrent collapse, and fails terminally. Fourier filtering separates circuit failure from masking. Across 43 checkpoints over five seeds and three regimes, the task-aligned family reaches exactly 100% alone. In circuit failure it no longer solves the task; in masking it remains perfect while the full model reaches 45.85%, giving a positive margin on every example, including errors, but being outvoted by a near-equal adversarial remainder. Rescaling it restores 99.9%; grokking is the same condition resolving upward. The task selects the family, swapping $(k,k)$ for $(k,-k)$ under subtraction. Across an abrupt collapse, standard Fourier support is unchanged and the power-distribution cosine remains 0.9899.

Gated Removal of Normalization in Transformers Enables Stable Training and Efficient Inference

Normalization is widely viewed as essential for stabilizing Transformer training. We revisit this assumption for pre-norm Transformers and ask to what extent sample-dependent normalization is needed inside Transformer blocks. We introduce TaperNorm, a drop-in replacement for RMSNorm/LayerNorm that behaves exactly like the standard normalizer early in training and then smoothly tapers to a learned sample-independent linear/affine map. A single global gate is held at $g{=}1$ during gate warmup, used to calibrate the scaling branch via EMAs, and then cosine-decayed to $g{=}0$, at which point per-token statistics vanish and the resulting fixed scalings can be folded into adjacent linear projections. Our theoretical and empirical results isolate scale anchoring as the key role played by output normalization: as a (near) $0$-homogeneous map it removes radial gradients at the output, whereas without such an anchor cross-entropy encourages unbounded logit growth (``logit chasing''). We further show that a simple fixed-target auxiliary loss on the pre-logit residual-stream scale provides an explicit alternative anchor and can aid removal of the final normalization layer. Empirically, TaperNorm matches normalized baselines under identical setups while eliminating per-token statistics and enabling these layers to be folded into adjacent linear projections at inference. On an efficiency microbenchmark, folding internal scalings yields up to $1.22\times$ higher throughput in last-token logits mode. These results take a step towards norm-free Transformers while identifying the special role output normalization plays.