paper: protocol-matched external benchmark positioning + 2024-25 related work

[ronaldtse]

Summary

• New paper results subsection Positioning against public benchmarks: verified SadeedDiac-25 leaderboard (teacher r6 2.5793 = best dedicated model, 2nd overall behind Claude-3.7-Sonnet's 1.3941, ahead of GLM-5.2 reproduction 2.6911 / Gemini-Flash 3.1926 / GPT-4 3.8645 / Sadeed-1.5B 7.2915), Hebrew Nakdimon-domain comparison (s45 16.58% DER vs DictaBERT-large 35.63% on the same test), Thai no-external-benchmark disclosure, Persian SentenceBench scoping
• Related work upgraded with the 2024-2025 landscape: SadeedDiac-25 (arXiv 2504.21635), multi-reference evaluation (Mohamed & Mubarak, EMNLP 2025), LLM-prompted G2P (arXiv 2409.08554, 8.30% PER on its own benchmark), intermediate-language G2P (arXiv 2505.06599), D-Nikud (arXiv 2402.00075), DiVRit (arXiv 2510.26521)
• Knowledge distillation section positions on-policy distillation / GKD (MiniLLM, Agarwal et al.) as the open lever against the measured flat-distribution pathology; added as explicit future work in Discussion
• Discussion adds benchmark-fragmentation limitation; notes teacher-tier beam gain is language-dependent (Hebrew +12 DER points, Arabic noise)
• RESULTS.md: Persian claim scoped to SentenceBench (concurrent lines use different PER benchmarks); ara-diac-small section gains leaderboard context; 9 new references, all verified against primary sources

Numbers cross-checked against the primary logs (rababa docs/RESULTS.md r6 verdict table; secryst-train arabic comparison table). No cross-protocol comparisons are claimed anywhere.

Test plan

• asciidoc table/footnote syntax matches existing sections
• every external number traced to a primary results log or a fetched citation page
• full-set ara-diac-small number lands separately and updates the subset value