Reproduction of arXiv 2512.19428: test perplexity gap 22.6% vs size-matched Transformer (242.94 vs 198.17), wider than paper’s claimed 10–15%. Custom CUDA kernels ~2× inference speedup reported.
Reproduction of arXiv 2512.19428: test perplexity gap 22.6% vs size-matched Transformer (242.94 vs 198.17), wider than paper’s claimed 10–15%. Custom CUDA kernels ~2× inference speedup reported.