{"id":"arxiv/2606.15036","url":"https://arxiv.org/abs/2606.15036","tags":["cs.LG","math.NT"],"score":0,"title":"Transformers Learn the Mestre-Nagao Heuristic","domain":"cs.LG","source":"arxiv","authors":["Pranav Venkata Konda"],"reviews":0,"subject":"cs.LG","summary":"We train a two-layer transformer encoder to classify rational elliptic curves $E/\\mathbb{Q}$ of conductor $\\leq 10000$ as either rank 0 or rank 1 from the first 128 normalized Frobenius traces. We achieve >99% accuracy on both classes, and accuracy is essentially unchanged on test curves with no isogeny or quadratic-twist relative in the training set. We then apply techniques from mechanistic interpretability such as attention analysis, linear probing, activation patching, logit attribution, and neuron-level circuit analysis to reverse-engineer the algorithm the (centroid in function space)…","category":"research","comments":0,"disputed":false,"division":"cs","mentions":0,"discussed":0,"published":"2026-06-13","reactions":{}}