Causal Whitespace Patching for Korean Byte-Latent Language Models: Quality-Preserving Latency and the Limits of Scale Amplification

Abstract

Byte-latent language models amortize global Transformer compute by grouping raw bytes into patches, but fewer patches need not preserve quality or reduce end-to-end latency. This study tests causal whitespace-informed patching on Korean in identical byte-latent graphs, separating boundary placement, patch rate, learned-router cost, and cached autoregressive time-to-output. At 19.6M parameters, the quality-matched whitespace policy reduced controlled end-to-end latency by 2.628% and strict-valid free-running latency by 2.531% relative to the C86 codepoint baseline. At 188.6M parameters, a quality-rescued W80 replication reduced controlled latency by 2.887% and free-running latency by 2.475%. The prespecified 10% target was not met, and model size did not amplify the percentage gain. The result is therefore a bounded positive systems finding and a negative scale-amplification finding, not a general 10% speedup or a production-ready Korean LLM. This open record contains the named preprint, self-contained arXiv source, curated reproducibility bundle, citation metadata, and SHA-256 checksums. It excludes raw corpora, model checkpoints, record identifiers, raw generated outputs, and per-sequence loss arrays. Code and reproducibility materials are Apache-2.0; the paper is CC BY 4.0.