
Wanna check how well a model can share knowledge between languages? Of course you do! 🤩 But can you do it without access to the model’s weights? Now you can with ECLeKTic 🤯

A postdoc at @CHIA_Cambridge and @CambridgeLTL. If I'm not here I'm at @omagolda following the news

Wanna check how well a model can share knowledge between languages? Of course you do! 🤩 But can you do it without access to the model’s weights? Now you can with ECLeKTic 🤯

Time understanding is an important faculty of human cognition. Is it the same for LLMs? According to benchmarks, reasoning about time can be broken down into understanding 1. calendars and clocks, which requires quantitative reasoning, and 2. the flow and logic of time, which needs logical reasoning. But we found that these benchmarks do not measure what we think they are measuring about time understanding in LLMs.

Our second keynote is by @JumeletJ , talking about multilingual evaluation at scale! #MRL26 #EMNLP2026

Next up is @mdlhx for our first keynote about how Multilingual NLP is English-centric and what can we do about it #MRL26 #EMNLP2026

Today @EjZhou582 and I presented our EMNLP paper about how independently trained monolingual language models converge on shared representations! aclanthology.org/2026.emnlp-… #EMNLP2026 #MultilingualNLP

Seats are starting to fill up for today's #MRL26 Workshop! Come join us for today's event! See you at Room F3. Follow this account for live tweets and updates! #EMNLP2026

As #EMNLP2026 wraps up, we warmly invite you to join us tomorrow at the MRL Workshop! 📆 09:00-17:00 GMT+1 🚶♂️In-person: Room F3 HUNGEXPO 💻 Virtual: underline.io/events/536/rece… Also, follow this account as we live-tweet the workshop tomorrow! #MRL26

Don't miss @annalkorhonen 's keynote and the remaining poster sessions tomorrow! And on Wednesday, @omerNLP and @ljvmiranda are co-organizing the @mrl_workshop ! Just drop by! Enjoy tonight's socials! 🎉
The @Cambridge_Uni Language Technology Lab is heading to Budapest for #EMNLP2026 next week! 🇭🇺 Our very own @annalkorhonen will deliver a keynote, "AI for the Many: Starting from the World, Not the Model" 🗓️ October 27, 14:00–15:00 Full sched below, come say hi! 👋

It turns out tokenizer evaluations might not tell us as much about how models turn out as we thought (or hoped). Very interesting results in our new EMNLP findings paper!
🆕 Findings of #EMNLP2026 🆕 Can you tell a good tokenizer from a bad one without training a model? We tested 29 intrinsic tokenizer metrics against 159 language models in nine languages. None of them reliably predicts downstream performance across suite and language combos 🧵

Will you be at EMNLP? Come check our lab's work: papers, a workshop and a keynote
The @Cambridge_Uni Language Technology Lab is heading to Budapest for #EMNLP2026 next week! 🇭🇺 Our very own @annalkorhonen will deliver a keynote, "AI for the Many: Starting from the World, Not the Model" 🗓️ October 27, 14:00–15:00 Full sched below, come say hi! 👋

Super excited to share that our method to convert language models to byte-level has been published in @Nature! This was originally the approach behind Bolmo. We now generalized it to other model families. All checkpoints (including new Qwen and Llama models) and data are public!
Our paper on retrofitting language models to operate over bytes – the approach behind Bolmo – has been accepted to Nature! 🎉 We’re also releasing new checkpoints that extend our method from Olmo to Qwen & Llama. 🧵 nature.com/articles/s41586-0…

We trained first language model for Yiddish! On the way we learned how low-resource tasks are still a challenge for modern LLMs. See us at a poster session at 11 am today @COLM_conf P.S. MameLoshne is rather BabeLoshn for me, as Yiddish was native lanaguage of my grandmother.
Yiddish combines Germanic grammar, Hebrew script, and fusion vocabulary. Multilingual LMs can produce readable Yiddish while missing its distinctive vocabulary and grammar. The result is a flatter version of the language.

very cool !
We trained an LLM for Yiddish! Existing pretraining data is only about 40% real Yiddish and models miss a lot of what makes the language what it is. MameLoshnLM fixes quite a bit of that! Come hear @UrikaUri talk about it at #COLM2026. Tmrw 11am Grand Ballroom Poster #123

We trained an LLM for Yiddish! Existing pretraining data is only about 40% real Yiddish and models miss a lot of what makes the language what it is. MameLoshnLM fixes quite a bit of that! Come hear @UrikaUri talk about it at #COLM2026. Tmrw 11am Grand Ballroom Poster #123
Yiddish combines Germanic grammar, Hebrew script, and fusion vocabulary. Multilingual LMs can produce readable Yiddish while missing its distinctive vocabulary and grammar. The result is a flatter version of the language.

It’s been a pleasure to contribute to this survey and to see it come together! @marco_computers did an amazing job leading this highly ambitious project which should become the go-to resource for any questions on or adjacent to tokenization.
🚨 [Token][ization] Paper Alert 🚨 Tokenization is a wildly understudied area of language modeling despite it having effects across all of NLP. Over the past ~8 months, 32 (!) tokenizer researchers put together the most comprehensive survey of the field. Check it out!

Big start of #TOKtober ! 🎃 👉 Just out: Token _ization Survey 👉 Oct 9: Tokenization Workhop @tokshop2025 @COLM_conf 👉 San Francisco 👉Oct 28: Tokenization tutorial @emnlpmeeting 👉Budapest Save the dates! Links below