MKAR starts with a small-scale reproduction of Attention Residuals. The first concern is matching the residual mechanism closely enough that comparisons have a clear baseline.
The next direction explores retrieval between separate autoregressive models. A synthetic image-and-text sequence task provides a controlled place to examine what information crosses between the paths.
The work is structured around a manageable training budget and explicit comparisons. Reproducing the mechanism and testing the proposed extension are separate questions. Neither the reproduction setup nor the extension design establishes an improvement by itself.