{"id":"arxiv/2607.17696","url":"https://arxiv.org/abs/2607.17696","tags":["stat.ML","cs.LG","math.OC"],"score":0,"title":"An Adjoint-Sensitivity Framework for Lost-in-the-Middle Phenomena in Causal Residual Transformers","domain":"stat.ML","source":"arxiv","authors":["Cheng Huan","Hongwei Yuan"],"reviews":0,"subject":"stat.ML","summary":"We develop an adjoint-sensitivity framework for positional influence in causal residual Transformers and separate unconditional analytic results from conditional boundary-shape conclusions. The principal unconditional theorem is the residual-to-depth-flow estimate for layer controls converging in $L^1$, complemented by a finite-token-to-Volterra attention estimate that explicitly controls the first cells near the causal endpoint. We define a normalized adjoint-energy influence density and derive its exact evolution along full-batch gradient flow. The adjoint admits an exact generator-term…","category":"research","comments":0,"disputed":false,"division":"stat","mentions":0,"discussed":0,"published":"2026-07-20","reactions":{}}