Hybrid transformer architecture replacing discrete layers with Neural ODE blocks, enabling continuous semantic control (sentiment, style) at inference time via learned steering signals

Neural ODE Transformer in Julia, trained via adjoint sensitivity methods. Matched-architecture vs. discrete Transformer on Penn Treebank: 119.9 val perplexity vs 113.7 discrete. Single-seed result; discrete wins at this scale as expected, continuous-depth advantage expected to emerge at larger scale.