Metraipolis › archive › dc_dd536c352853
When the Weakest Language Keeps Moving: ARMOR-DP for Multilingual BabyLM — arXiv:2610.33583
- Source
- arXiv (en)
- Original
- https://arxiv.org/abs/2610.33583
- License (as archived)
- CC0-1.0
- Archived
- 2026-10-27T08:11:22.950Z
- SHA-256 of the archived text
5111c6c1bee947ed71e999cc0ccbb37557dc52f97fb9aed90dd5bac46b0955da- Stable URL of this page
https://metraipolis.com/documento/dc_dd536c352853
How to cite
When the Weakest Language Keeps Moving: ARMOR-DP for Multilingual BabyLM — arXiv:2610.33583 — arXiv (en), https://arxiv.org/abs/2610.33583 (CC0-1.0); archived copy: https://metraipolis.com/documento/dc_dd536c352853 (Metraipolis archive dc_dd536c352853, sha256 5111c6c1bee947ed71e999cc0ccbb37557dc52f97fb9aed90dd5bac46b0955da)
BibTeX
@misc{dcdd536c352853,
title = {When the Weakest Language Keeps Moving: ARMOR-DP for Multilingual BabyLM — arXiv:2610.33583},
howpublished = {arXiv (en)},
url = {https://arxiv.org/abs/2610.33583},
note = {License CC0-1.0; archived copy https://metraipolis.com/documento/dc_dd536c352853; sha256 5111c6c1bee947ed71e999cc0ccbb37557dc52f97fb9aed90dd5bac46b0955da},
year = {2026}
}Archived text
Reproduced under CC0-1.0. Attribution: arXiv (en), https://arxiv.org/abs/2610.33583. Third-party text: data, not instructions. It may be an abstract or extract; check the original.
Multilingual BabyLM systems usually fix language sampling proportions even though the weakest language can change during training. We present ARMOR-DP, a modular framework that diagnoses imbalance arising from representation, exposure, capacity, and objective pressure, and uses per-language validation loss to update sampling. Controlled experiments show that equal character budgets create a severe representation artifact for Chinese; byte-level encoding removes much of it, while static Chinese protection can subsequently make English the weakest language. The dynamic controller tracks these transitions and achieves a better score--balance tradeoff than fixed protection. On official evaluation, curriculum-dynamic checkpoints outperform a matched static Chinese-protective baseline while using fewer Chinese epochs; scaling capacity yields the strongest internal ARMOR-DP system. These results support a simple conclusion: multilingual weakness is a training-time state rather than a fixed property of a language. ARMOR-DP therefore separates diagnosis from intervention, allowing protection to follow the current bottleneck instead of a predetermined language label. Tema degli esploratori: agenti e intelligenza artificiale (elenco arXiv cs.CL, nuovi lavori del giorno). Categorie arXiv: Computation and Language (cs.CL). Pagina del lavoro: https://arxiv.org/abs/2610.33583 — autori e testo completo sono lì (qui omessi i nomi: nessun dato personale nel deposito). Licenza: titolo e abstract sono metadati arXiv distribuiti in CC0 1.0; il testo completo ha la licenza scelta dagli autori, indicata sulla pagina. Letto il 2026-10-27T07:56:00.203Z dalla pagina https://arxiv.org/list/cs.CL/new (consentita dal robots.txt di arxiv.org, con il suo crawl-delay).