Escape Meridian

Chelombitko, Iaroslav, Hämäläinen, Mika · Nov 14, 2026

Do Rule-Based Tools Help LLMs Translate Endangered Uralic Languages in the Prompt or over MCP?

Markdown · JSON · Original source

Authors Iaroslav Chelombitko and Mika Hämäläinen ask whether rule-based resources improve machine translation from Komi-Zyrian, Moksha, Erzya and Udmurt into English, and whether models benefit more from analyses supplied in the prompt or tools they call through the Model Context Protocol (MCP). In a manual evaluation, an expert linguist judged every English output from the strongest model across three conditions. Accepted translations rose from 13 to 43 out of 272 when analyses were put in the prompt, and to 55 with tools attached. The authors attribute the gain to the bilingual dictionary: for Udmurt, whose dictionary has no English side, the prompt condition brought no gain; for the other languages, an English gloss reached the prompt for 74 to 80% of source words. The comparison also exposes a practical hurdle: was harder to get working than . Of 31 , 14 never called a tool, usually because they could not produce a usable call, and half of calls made named an unsupported language. The authors release the data set, model outputs and tool-call logs. The abstract does not give enough detail to assess how these results vary across models or translation settings.

Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.

arXiv cs.AIArticle · arxiv.orgOpen original source

Original source

cs.AI · Article

Source details

Original source date: Nov 14, 2026 (2026-11-14).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-17T11:16:46.380Z. First observed: 2026-11-17T08:30:57.271Z. Retrieved: 2026-11-17T11:05:39.289Z.

Referenced by

Recorded citations in public readings. A reference alone does not establish agreement or influence.

No public references recorded here yet.

Reference data (JSON)

Limits

The abstract reports manual judgments by an expert linguist but does not provide the evaluation criteria or further methodological detail. The results described are for the strongest model under three conditions; the abstract does not establish how results generalize across other models or translation settings.

Open question

Would the results hold across models and translation settings beyond the strongest model evaluated here?

Source-stated

The authors compare prompt-supplied analyses with model-called tools for translating four endangered Uralic languages into English.

We ask whether giving a model access to these resources improves machine translation from the four languages into English, and whether it matters how the access is provided

Source-stated

In manual evaluation of 816 outputs across three conditions, accepted translations rose from 13 to 43 out of 272 with prompt-supplied analyses, and to 55 with tools attached.

With the analyses in the prompt, the number of accepted translations rises from 13 to 43 out of 272; with the tools attached it rises to 55.

Source-stated

The authors say the gains came from the bilingual dictionary; Udmurt's dictionary lacks an English side and its prompt-based condition brought no gain.

The gain comes from the bilingual dictionary. For Udmurt, whose dictionary has no English side, the prompt-based condition brings no gain at all

Source-stated

Tool use was unreliable among the 31 open-weight models tested: 14 never called a tool, and half of calls made named an unsupported language.

Of 31 open-weight models, 14 never call a tool, 11 of them because they cannot produce a usable call, and half of the calls that are made name a language the toolkit does not support.

Source-stated

The authors state that they release the data set, model outputs and tool-call logs.

We release the data set, every model output and all tool-call logs.

Source and contributor context

Chelombitko, Iaroslav, Hämäläinen, Mika · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Large language model

A language model with many learned numerical settings, used to process and generate language.

Tool use

A model's ability to operate other software, such as a search engine, a calculator or a code runner, while it works.

Prompt injection

Hidden instructions placed in a document, email or web page that try to take control of an AI system reading it.

Open weights

A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

Editorial provenance

Version 1. Publication ID: 240cb954-61aa-408c-8f77-5eff57263087.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-17T11:16:46.380Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 938f2c27a90cc6d8a2088db3ebe0466980a972d7ad86c8355416a1d81820807c. Source text hash: 88b20186373670ead76fa324ec50625240efec1912b2835ace2aac52042f493e.

Revision and source-check history

Version 1 · 2026-11-17T11:16:46.380Z

Do Rule-Based Tools Help LLMs Translate Endangered Uralic Languages in the Prompt or over MCP?

Authors Iaroslav Chelombitko and Mika Hämäläinen ask whether rule-based resources improve machine translation from Komi-Zyrian, Moksha, Erzya and Udmurt into English, and whether models benefit more from analyses supplied in the prompt or tools they call through the Model Context Protocol (MCP). In a manual evaluation, an expert linguist judged every English output from the strongest model across three conditions. Accepted translations rose from 13 to 43 out of 272 when analyses were put in the prompt, and to 55 with tools attached. The authors attribute the gain to the bilingual dictionary: for Udmurt, whose dictionary has no English side, the prompt condition brought no gain; for the other languages, an English gloss reached the prompt for 74 to 80% of source words. The comparison also exposes a practical hurdle: was harder to get working than . Of 31 , 14 never called a tool, usually because they could not produce a usable call, and half of calls made named an unsupported language. The authors release the data set, model outputs and tool-call logs. The abstract does not give enough detail to assess how these results vary across models or translation settings.

Original source

cs.AI · Article

Source details

Original source date: Nov 14, 2026 (2026-11-14).

Date basis: Date basis not provided.

Added to Escape Meridian: 2026-11-17T11:16:46.380Z. First observed: 2026-11-17T08:30:57.271Z. Retrieved: 2026-11-17T11:05:39.289Z.

Limits

The abstract reports manual judgments by an expert linguist but does not provide the evaluation criteria or further methodological detail. The results described are for the strongest model under three conditions; the abstract does not establish how results generalize across other models or translation settings.

Open question

Would the results hold across models and translation settings beyond the strongest model evaluated here?

Source-stated

The authors compare prompt-supplied analyses with model-called tools for translating four endangered Uralic languages into English.

We ask whether giving a model access to these resources improves machine translation from the four languages into English, and whether it matters how the access is provided

Source-stated

In manual evaluation of 816 outputs across three conditions, accepted translations rose from 13 to 43 out of 272 with prompt-supplied analyses, and to 55 with tools attached.

With the analyses in the prompt, the number of accepted translations rises from 13 to 43 out of 272; with the tools attached it rises to 55.

Source-stated

The authors say the gains came from the bilingual dictionary; Udmurt's dictionary lacks an English side and its prompt-based condition brought no gain.

The gain comes from the bilingual dictionary. For Udmurt, whose dictionary has no English side, the prompt-based condition brings no gain at all

Source-stated

Tool use was unreliable among the 31 open-weight models tested: 14 never called a tool, and half of calls made named an unsupported language.

Of 31 open-weight models, 14 never call a tool, 11 of them because they cannot produce a usable call, and half of the calls that are made name a language the toolkit does not support.

Source-stated

The authors state that they release the data set, model outputs and tool-call logs.

We release the data set, every model output and all tool-call logs.

Source and contributor context

Chelombitko, Iaroslav, Hämäläinen, Mika · Article author

These are credits stated by the source; they do not establish a profile match or authorship of every passage.

Source media

No suitable media was found by the supported page extraction.

Terms and references

Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.

Large language model

A language model with many learned numerical settings, used to process and generate language.

Tool use

A model's ability to operate other software, such as a search engine, a calculator or a code runner, while it works.

Prompt injection

Hidden instructions placed in a document, email or web page that try to take control of an AI system reading it.

Open weights

A model whose trained numbers are published so others can download, run and inspect it; the license and training data may still be restricted.

arXiv

arXiv is a research manuscript repository where authors share papers that may not yet have been peer reviewed.

Editorial provenance

Version 1. Publication ID: 240cb954-61aa-408c-8f77-5eff57263087.

Prepared by gpt-6-luna; grounding review by gpt-6-luna.

Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-17T11:16:46.380Z. This checks fidelity to the source, not whether its claims are independently true.

Editorial selection: accept; reasons: material_development.

Observation: 938f2c27a90cc6d8a2088db3ebe0466980a972d7ad86c8355416a1d81820807c. Source text hash: 88b20186373670ead76fa324ec50625240efec1912b2835ace2aac52042f493e.

uncertain · 2026-11-17T11:16:46.380Z · anthropic/claude-opus-5.5

Publication: 240cb954-61aa-408c-8f77-5eff57263087

attribution — pass: The reading correctly names Iaroslav Chelombitko and Mika Hämäläinen as the authors. It keeps the findings attributed to their arXiv paper, including the dictionary explanation ('The authors attribute the gain').

Authors:Iaroslav Chelombitko, Mika Hämäläinen

summary — uncertain: The figures are faithful, but a non-specialist may not follow the central terms. 'Rule-based resources' is never explained, and the reading drops the abstract's examples (finite-state analysers, constraint grammars, bilingual dictionaries). 'Open-weight models', 'English gloss' and 'prompt injection' are also unexplained; the last can be misread as the security attack, though here it means writing analyses into the prompt. The reading also never says why the result matters: these languages have mature rule-based resources that models do not use. Smallest repair: name the three resource types and that language models do not use them, and add plain glosses or tooltips for open-weight, gloss and prompt injection.

among them finite-state analysers, constraint grammars and bilingual dictionaries, which language models do not use

date — pass: The originalPublished value of 2026-11-14 matches the v1 submission date shown on the arXiv page. The feed date is later.

[Submitted on 14 Nov 2026]

support — pass: Each claim matches the abstract. This covers the 13, 43 and 55 out of 272 accepted translations, the 816 sentences, the dictionary explanation and the Udmurt exception, the 74 to 80% gloss coverage, and the 14 of 31 models that never called a tool (11 unable to produce a usable call). It also covers the release of the data set, outputs and logs. 'Usually' fairly renders 11 of 14. The question and limitations are framed as open issues, not claims.

Of 31 open-weight models, 14 never call a tool, 11 of them because they cannot produce a usable call

media — pass: The selected media is the verified arXiv abstract page for this paper. It is the direct original source.

Challenge or correct this reading

Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.