Slicing a Nested Byte-Level Vocabulary Is Bit-Exact and Cuts 66% of Weights, But the Shared Model Loses to Every Specialist
arXiv 2608.28151 (2026-08-28, cs.AI/cs.CL/cs.IR) pre-registers five claims with margins, seeds, contrasts and a stop rule, then trains 30 models at 3.1M and 10.6M parameters on 200M tokens each. Slicing a prefix-nested BPE vocabulary is numerically exact, reproducing the restricted full model's logits bit for bit across 76 checks and removing 66% of deployed weights without changing latency, but the shared model trails a fixed-cap specialist by 3.64% bits per byte at 32k against a 1% margin and 2.96% at 8k against a 2% margin. A 2x2 ablation shows the control token contributes only +0.07% to +0.13% with intervals crossing zero while output restriction costs +0.47% to +1.19%, and the robustness gain under typographical noise (12.5 to 15.4 points less degradation) traces to multi-granularity training rather than the conditioning mechanism.
↳ Follow the thread