Exploring Multilingual and Local-Language Encoding Extensions to LIMIT

[Arsh123344423]

I’m interested in extending the LIMIT benchmark to explore how multilingual and local-language encoding affects the observed theoretical limitations of single-vector embedding-based retrieval. While LIMIT convincingly demonstrates that these limitations are language-agnostic and dimension-dependent, it would be valuable to empirically analyze whether encoding queries and documents in their native or local languages changes the failure structure, retrieval dynamics, or error distribution across languages.
Specifically, I’d like to investigate
(1) multilingual variants of the LIMIT dataset,
(2) language-conditioned or language-aware embedding pipelines,
(3) cross-lingual versus mono-lingual retrieval settings, while keeping the theoretical guarantees intact.
My goal is not to circumvent the LIMIT result, but to better understand how these theoretical constraints manifest in multilingual conversational and retrieval systems. I’d appreciate feedback on whether such an extension aligns with the project’s direction and any guidance on best practices for contributing this experimentally.

[orionw]

Hi @Arsh123344423! Thanks for the interest.

You're certainly welcome to create a new repo and extend LIMIT to other languages. You might find empirical differences that could be interesting.

However, our theoretical results are independent of language (as shown in the free embedding experiments where the inputs are just ids).

[Arsh123344423]

Yes absolutely!! @orionw so am i allowed to create a fork for LIMIT over this idea

[Arsh123344423]

@orionw,Thanks for clarifying, that makes sense. just a little confirmation

I’ll proceed by creating a separate companion repository that extends LIMIT empirically to multilingual and cross-lingual settings, while keeping the original theoretical guarantees unchanged.

I’ll make it explicit that the goal is not to challenge the language-agnostic LIMIT result, but to study how those same geometric constraints manifest under multilingual encoding and alignment.

I’ll share results back here once there’s something concrete. Thanks again for the guidance.

And so sorry for disturbing you again