{"id":"arxiv/2611.00585","url":"https://arxiv.org/abs/2611.00585","tags":["cs.LG","cs.AI","cs.CL","math.OC"],"score":0,"title":"Conic Certification of Transformer Attention","domain":"cs.LG","source":"arxiv","authors":["Masoud Ataei"],"reviews":0,"subject":"cs.LG","summary":"Attention is the computational core of large language models, yet the verifiers that bound how far an attention head's output can move under input perturbations give no guarantee on how loose their bounds are. A weak certificate therefore cannot be told apart from a fragile model. We develop a conic programming theory of attention certification that supplies this guarantee. We prove that the difficulty lies not in the softmax but in the bilinear interaction of a varying query with the keys. This interaction makes certification NP-hard over boxes and inapproximable once the attention logits…","category":"research","comments":0,"disputed":false,"division":"cs","mentions":0,"discussed":0,"published":"2026-10-31","reactions":{}}