Exact Attention Sensitivity and the Geometry of Transformer Stability
Fuente:
arXiv
Salvato in:
| Autore principale: | |
|---|---|
| Natura: | Preprint |
| Pubblicazione: |
2026
|
| Soggetti: | |
| Accesso online: | |
| Tags: |
Aggiungi Tag
Nessun Tag, puoi essere il primo ad aggiungerne!!
|
| _version_ | 1866912917521170432 |
|---|---|
| author | Emadi, Seyed Morteza |
| author_facet | Emadi, Seyed Morteza |
| contents | Despite powering modern AI, transformers remain mysteriously brittle to train. We develop a stability theory that explains why pre-LayerNorm works, why DeepNorm uses $N^{-1/4}$ scaling, and why warmup is necessary, all from first principles. Our framework has two pillars: (1) We derive the \emph{exact} operator norm of the softmax Jacobian, $\|J_{softmax}(u/τ)\|_{\infty\to 1} = θ(p)/τ$, where the balanced-mass factor $θ(p)\in[0,1]$ quantifies attention sensitivity. (2) We introduce a block-$\infty$/RMS geometry aligned with tokenwise computation, yielding Lipschitz bounds independent of sequence length. Using this framework, we prove that pre-LN preserves identity gradient paths while post-LN compounds LayerNorm Jacobians exponentially with depth, and we show that DeepNorm's $N^{-1/4}$ emerges from the quartic structure of attention's four projection matrices. We validate our theory on 774M-parameter models and find that, contrary to the intuition that attention sharpens during training to reduce sensitivity, $θ(p) \approx 1$ persists throughout. Transformer stability arises entirely from architectural gradient flow, not from attention dynamics. This finding changes how we reason about training: the architecture itself must handle sensitivity, not learned attention patterns. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2602_18849 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Exact Attention Sensitivity and the Geometry of Transformer Stability Emadi, Seyed Morteza Machine Learning Artificial Intelligence Despite powering modern AI, transformers remain mysteriously brittle to train. We develop a stability theory that explains why pre-LayerNorm works, why DeepNorm uses $N^{-1/4}$ scaling, and why warmup is necessary, all from first principles. Our framework has two pillars: (1) We derive the \emph{exact} operator norm of the softmax Jacobian, $\|J_{softmax}(u/τ)\|_{\infty\to 1} = θ(p)/τ$, where the balanced-mass factor $θ(p)\in[0,1]$ quantifies attention sensitivity. (2) We introduce a block-$\infty$/RMS geometry aligned with tokenwise computation, yielding Lipschitz bounds independent of sequence length. Using this framework, we prove that pre-LN preserves identity gradient paths while post-LN compounds LayerNorm Jacobians exponentially with depth, and we show that DeepNorm's $N^{-1/4}$ emerges from the quartic structure of attention's four projection matrices. We validate our theory on 774M-parameter models and find that, contrary to the intuition that attention sharpens during training to reduce sensitivity, $θ(p) \approx 1$ persists throughout. Transformer stability arises entirely from architectural gradient flow, not from attention dynamics. This finding changes how we reason about training: the architecture itself must handle sensitivity, not learned attention patterns. |
| title | Exact Attention Sensitivity and the Geometry of Transformer Stability |
| topic | Machine Learning Artificial Intelligence |
| url | https://arxiv.org/abs/2602.18849 |