Reproduction: Incremental BPE Tokenization

Jiang & Gong, ICML 2026 (OpenReview ZbWgrDzCQo · arXiv 2605.30813) — independent Python reimplementation, evidence-first, every number from a committed protocol-compliant run

Claim 1 · Monotonic Path Property

Reproduced (empirically)

Zero violations across exhaustive candidate×history cross-products, per-byte upward-closure checks, 200 random dictionaries, and real 50k-vocab corpus runs.

Claim 2 · Algorithm design

Partially reproduced

Composition verified end-to-end at real scale. Missing §5.3 / §6.2 mechanisms: their absence measured (tail latency linear in depth; eager 34× vs paper's ~10%).

Claim 3 · 3.13× drop-in speedup

Equivalence reproduced · speedup untestable

Token-for-token identical to HF tokenizers over the paper's English corpus recipe. The speedup sits behind a measured 52× Python-vs-Rust constant — pre-declared scope limit.

Claim 4 · Pathological robustness

Reproduced (era-appropriate baseline)

Ours flat (slope 0.99); tiktoken 0.8.0 quadratic (2.04–2.08, incl. CL100K). Finding: current tiktoken fixed the quadratic upstream (1.14–1.16).

0.987ours, log-log slope, real vocab ('a'*n)
2.041tiktoken 0.8.0 on CL100K (quadratic)
114,399tokens, zero divergences vs HF tokenizers
5.7×our Python beats old Rust tiktoken @200 kB pathological
44 / 2,126tests / subtests, green in CI
Evidence & code: github.com/NabarunCode/icml-2026-reproductions · per-file independence record (PROVENANCE.md) · hash-pinned data · clean-tree environment records on every run