Lightnews — Scholar-powered news

Reposted by Tom Kocmi

Julia Kreutzer @juliakreutzer.bsky.social · 15h

Ready for our poster today at #COLM2025!

💭This paper has had an interesting journey, come find out and discuss with us! @swetaagrawal.bsky.social @kocmitom.bsky.social

Side note: being a parent in research does have its perks, poster transportation solved ✅

1 8

Tom Kocmi @kocmitom.bsky.social · 22d

This project wouldn’t have been possible without the brilliant minds driving the work: Lorenzo Proietti, @sted19.bsky.social and @zouharvi.bsky.social

3

Tom Kocmi @kocmitom.bsky.social · 22d

One way to raise the bar is by rethinking the source selection process: instead of random samples, we built model that chooses the most difficult data for translation. And we’ve already put our work into practice: this year’s WMT25 General MT test set use our approach to make eval more challenging.

1 2

Tom Kocmi @kocmitom.bsky.social · 22d

🚩Machine Translation is far from “solved” - the test sets just got too easy. 🚩

Yes, the systems are much stronger. But the other half of the story is that test sets haven’t kept up. It’s no longer enough to just take a random news article and expect systems to stumble.

Stefano @sted19.bsky.social · 22d

Our new #EMNLP2025 paper is out: "Estimating Machine Translation Difficulty"! 🚀

Are today's #MachineTranslation systems flawless? When SOTA models all achieve near-perfect scores on standard benchmarks, we hit an evaluation ceiling. How can we tell their true capabilities and drive future progress?

1 1 6

Tom Kocmi @kocmitom.bsky.social · Aug 28

Oh, and the best part: we’re releasing the weights so researchers can run wild with it. Stay tuned for our upcoming technical report!

cohere.com/blog/command...

Command A Translate: Secure translation for global enterprises

The new industry standard for secure, enterprise-ready machine translation.

cohere.com

Tom Kocmi @kocmitom.bsky.social · Aug 28

🚀 Thrilled to share what I’ve been working on at Cohere!

What began in January as a scribble in my notebook “how challenging would it be...” turned into a fully-fledged translation model that outperforms both open and closed-source systems, including long-standing MT leaders.

1 1 5

Tom Kocmi @kocmitom.bsky.social · Aug 26

A correction: we obtained 22 multilingual systems while in contrast we got only 14 bilingual systems, highlighting a shift in the field towards multilinguality.

Tom Kocmi @kocmitom.bsky.social · Aug 23

We received 14 specialized systems while 10 are multilingual. And almost all participants finetuned some LLMs.

In contrast to previous years, constrained systems are now reaching top-tier rankings, challenging the dominance of unconstrained ones.

Stay tuned for the 20th anniversary WMT conference.

1 1

Tom Kocmi @kocmitom.bsky.social · Aug 23

We saw increased momentum in participation growth this year: 36 unique teams competing to improve the performance of MT. Furthermore, we added collected outputs of 24 popular LLMs and online systems. Reaching 50 evaluated systems in our annual benchmark.

1 1 3

Tom Kocmi @kocmitom.bsky.social · Aug 23

📊 Preliminary ranking of WMT 2025 General Machine Translation benchmark is here!

But don't draw conclusions just yet - automatic metrics are biased for techniques like metric as a reward model or MBR. The official human ranking will be part of General MT findings at WMT.

arxiv.org/abs/2508.14909

Preliminary Ranking of WMT25 General Machine Translation Systems

We present the preliminary ranking of the WMT25 General Machine Translation Shared Task, in which MT systems have been evaluated using automatic metrics. As this ranking is based on automatic evaluati...

arxiv.org

1 4 9

Tom Kocmi @kocmitom.bsky.social · Jun 26

Hey, hey! 🎉 We’ve released the blind test set for this year’s WMT General MT and multilingual instruction tasks. Submit your systems to the special 20th anniversary of the conference and see how you compare to others!
The deadline is next week on 3rd July.
www2.statmt.org/wmt25/

WMT 2025

www2.statmt.org

1

Tom Kocmi @kocmitom.bsky.social · Apr 17

Tired of messy non-replicable multilingual LLM evaluation? So were we.

In our new paper, we experimentally illustrate common eval. issues and present how structured evaluation design, transparent reporting, and meta-evaluation can help us to build stronger models.

Julia Kreutzer @juliakreutzer.bsky.social · Apr 17

📖New preprint with Eleftheria Briakou @swetaagrawal.bsky.social @mziizm.bsky.social @kocmitom.bsky.social!

arxiv.org/abs/2504.11829

🌍It reflects experiences from my personal research journey: coming from MT into multilingual LLM research I missed reliable evaluations and evaluation research…

Screenshot of the paper header with title and author list and affiliations

1 7

Tom Kocmi @kocmitom.bsky.social · Mar 28

☀️ Summer internship at Cohere!
Are you excited about multilingual evaluation, human judgment, or meta-eval? Come help us explore how a rigorous eval really looks like while questioning the status quo in LLM evaluation.
I’m looking for an intern (EU timezone preferred), are you interested? Ping me!

2 2 7

Tom Kocmi @kocmitom.bsky.social · Mar 27

It’s here! Our new model’s technical report is out. I'm especially proud of the work we did on its multilingual capabilities - this was a massive, collective effort!

seraphinagt.bsky.social @seraphinagt.bsky.social · Mar 27

Today (two weeks after model launch 🔥) we're releasing a technical report of how we made Command A and R7B 🚀! It has detailed breakdowns of our training process, and evaluations per capability (tools, multilingual, code, reasoning, safety, enterprise, long context)🧵 1/3.

1

Tom Kocmi @kocmitom.bsky.social · Mar 11

Big news from WMT! 🎉 We are expanding beyond MT and launching a new multilingual instruction shared task. Our goal is to foster truly multilingual LLM evaluation and best practices in automatic and human evaluation. Join us and build the winning multilingual system!
www2.statmt.org/wmt25/multil...

Multilingual Instruction Shared Task

www2.statmt.org

1 7 12

Tom Kocmi @kocmitom.bsky.social · Mar 4

AI is evolving fast, and Aya Vision is proof of that. This open-weights model is designed to make LLM more powerful across languages and modalities, especially vision! Can’t wait to see the real-world applications, perhaps at WMT this year 😇

Cohere Labs @cohereforai.bsky.social · Mar 4

We hope Aya enables researchers and developers throughout the world to build upon this technology, ask deeper questions about multilingual AI, and develop tools that can support their communities.

Learn more:

cohere.com/blog/aya-vis...

Aya Vision: Expanding the Worlds AI Can See

Our state-of-the-art open-weights vision model offers a foundation for AI-enabled multilingual and multimodal communication globally. Today, Cohere For AI, Cohere’s open research arm, is proud to an...

cohere.com

2

Tom Kocmi @kocmitom.bsky.social · Mar 1

Huge shoutout to colleagues at Google & Unbabel for extending our WMT24 testset to 55 languages in four domains, this is game changer! 🚀

I really hope it puts the final nail in the coffin of FLORES or WMT14. The field is evolving, legacy testsets can't show your progress

arxiv.org/abs/2502.124...

WMT24++: Expanding the Language Coverage of WMT24 to 55 Languages & Dialects

As large language models (LLM) become more and more capable in languages other than English, it is important to collect benchmark datasets in order to evaluate their multilingual performance, includin...

arxiv.org

6 14

Tom Kocmi @kocmitom.bsky.social · Feb 20

* Revamped constrained track – No restrictions on training data except licensing; all open models under 20B parameters are allowed.

* More challenging sources; long-context translation; prompt preambles; and much more.

📌 All details are available at www2.statmt.org/wmt25/transl...

Shared Task: General Machine Translation

www2.statmt.org

2

Tom Kocmi @kocmitom.bsky.social · Feb 20

* New human-evaluated language pairs: EN–Arabic, EN–Estonian, EN–Korean, EN–Serbian, Czech–German, Bhojpuri–EN, Maasai–EN

* New multilingual subtask – Can you build a system that translates 30 languages?

* New modalities – Additional context from video and image (text-to-text remains the core).

1 4

Tom Kocmi @kocmitom.bsky.social · Feb 20

Guess what? The jubilee 🎉 20th iteration of WMT General MT 🎉 is here, and we want you to participate - as the entry barrier to make an impact is so low!

This isn’t just any repeat. We’ve kept what worked, removed what was outdated, and introduced many exciting new twists! Among the key changes are:

1 5 18

Tom Kocmi @kocmitom.bsky.social · Feb 9

Yeah, I haven't wrote a paper since it's just a different prompt. It's published in the github repository of GEMBA

1

Tom Kocmi @kocmitom.bsky.social · Feb 8

That one is extremely large, but we haven't used it either in the automatic ranking. Unfortunately I'm not aware of any API service for metrics

1

Tom Kocmi @kocmitom.bsky.social · Nov 20

🙏 A huge thank you to all organizers, partners, and participants for making this year's WMT General MT Shared Task a success! Stay tuned for WMT25 - many exciting changes are coming! 🎉

2

Tom Kocmi @kocmitom.bsky.social · Nov 20

🏆 Highlights from top systems:
✅ IOL-Research: led in constrained/open, winning 10/11 in its category.
✅ Unbabel-Tower70B: Best participant, winning 8/11 pairs.
✅ Claude-3.5-Sonnet: Best overall with 9/11 wins.
✅ Shoutout to Dubformer (speech) & CUNI-MH (strong constrained)

1 4

Tom Kocmi @kocmitom.bsky.social · Nov 20

📊 We introduced new robust and efficient human evaluation protocol: Error Span Annotations (ESA).
📄 Test sets are now finally document-level!
🌍 We've added three new language pairs, including English-Spanish where translations are near-perfect.
For more details, read our findings paper.

1