Cohere Labs

What an incredible week it’s been at #NeurIPS2025! 🎉

Today is our last one at the booth. We've had a great week connecting with our community in San Diego.

Join our community to continue to connect with our research team: https://cohere.com/research/open-science/application

What's the story of your legend?

Join ML researchers building their legends with 40 cards that capture our shared journey—explore and build yours: https://lab-legends.vercel.app/ 🎯

Just 1 day left until #NeurIPS2025 kicks off! The Cohere and Cohere Labs teams are ready to dive into a packed week of research, conversations, and community at the San Diego Convention Center✨

Come visit our booth — we’d love to chat and send you home with some swag!

... @markusfreitag.bsky.social, Roman Grundkiewicz, @yupenghou.bsky.social, @phikoehn.bsky.social, @juliakreutzer.bsky.social, Saab Mansour, @sted19.bsky.social, Lorenzo Proietti, Parker Riley, Eduardo Sánchez, @patuchen.bsky.social, Mariya Shmatova, @zouharvi.bsky.social

You can find all details in our paper www2.statmt.org/wmt25/pdf/20... or discuss with us next week at the WMT Conference at #EMNLP2025.

Led by @kocmitom.bsky.social, Ekaterina Artemova, Eleftherios Avramidis, Eleftheria Briakou, @pinzhen.bsky.social, @mziizm.bsky.social...

⚖️ LLM-as-a-judge: mixed reliability.

Top systems reach ~95% pairwise accuracy open-ended and summarization tasks.
Smaller ones barely beat coin-flip territory at ~55%.

🤖Naturalness is still a significant challenge.

Across open-ended generation and cross lingual summarization, the biggest weakness isn’t coherence or accuracy, but it is sounding like a native speaker. Many outputs still feel robotic or translated.

🧠English isn’t always easiest.

Models like Gemini 2.5 Pro and Claude 4 sometimes did better in Korean, German, or Spanish than in English when solving reasoning tasks.

🧩Linguistic reasoning remains the toughest nut. 🥥

Even top models scored below 50% on linguistic reasoning tasks, showing that structured linguistic deduction is still an open challenge.

🌐 Language coverage matters.

Models don’t support all languages equally, and this skews rankings. Smaller open models especially struggle with broad coverage, affecting their aggregate ranking ⚠️

🧩 Linguistic reasoning on unseen languages
📝 Open-ended generation testing naturalness and usefulness
📘 Cross-lingual summarization
🔁 Machine translation
🧑‍⚖️ LLM-as-a-Judge evaluating outputs of other models

All backed by human evals and public releases of data + outputs!
github.com/wmt-conferen...

How well do LLMs handle multilinguality? 🌍🤖

🔬We brought the rigor from Machine Translation evaluation to multilingual LLM benchmarking and organized the WMT25 Multilingual Instruction Shared Task spanning 30 languages and 5 subtasks.

River, Yinhong and I will all be in person and we look forward to the discussions!

Cohere Labs x EMNLP 2025: "When Life Gives You Samples: The Benefits of Scaling up Inference Compute for Multilingual LLMs"

Congrats to authors Ammar Khairi, Daniel D'souza, Ye Shen, @juliakreutzer.bsky.social, @sarahooker.bsky.social

📜 arxiv.org/abs/2506.20544

Cohere Labs x EMNLP 2025 "When Personalization Meets Reality: A Multi-Faceted Analysis of Personalized Preference Learning"

Congrats to authors Yijiang River Dong, @tiancheng.bsky.social, Yinhong Liu, Ahmet Üstün, Nigel Collier.

📜 arxiv.org/abs/2502.19158

Cohere Labs x EMNLP 2025: "The State of Multilingual LLM Safety Research: From Measuring The Language Gap To Mitigating It"

Congrats to authors @yongzx.bsky.social , Beyza Ermis, @mziizm.bsky.social, Stephen Bach, @juliakreutzer.bsky.social.

📜 arxiv.org/abs/2505.24119

Cohere Labs x EMNLP 2025: "Nexus: Adaptive Upcycling to Efficiently Pretrain Mixture of Experts"

Congrats to authors Nikolas Gritsch, Qizhen Zhang, @acyrl.bsky.social, @sarahooker.bsky.social and Ahmet Üstün.

📜 arxiv.org/abs/2408.15901

We’re thrilled to announce that some of our research will be presented at @emnlpmeeting.bsky.social next week! 🥳

If you’re attending the conference, don’t miss the chance to explore our work and connect with our team.

We're excited to hear from speakers including Ivan Zhang, Joelle Pineau, Marzieh Fadaee, Shayne Longpre and 20+ other presenters who will share insights on open science, collaborative research, and community-driven innovation.

Learn more and register now: https://tinyurl.com/CohereLabsConnect

Join us for inspiring keynotes, lightning talks, and interactive sessions that bring together curious minds from around the world. Throughout the conference, we’ll:

🔬 Showcase cutting-edge research
💡 Highlight meaningful collaborations
🤝 Inspire new partnerships

“Individually, we are one drop. Together, we are an ocean.” - Ryunosuke Satoro ✨

Cohere Labs is excited to announce Connect - a 3-day virtual conference celebrating the power of collaboration in open science!

📜Paper link: arxiv.org/pdf/2510.19806

Led by: David Mora, Viraat Aryabumi, @weiyinko-ml.bsky.social, @sarahooker.bsky.social, @juliakreutzer.bsky.social, and
@mziizm.bsky.social.

With this work we take a step toward principled approaches to multilingual synthetic data generation—an essential direction for developing adaptive, culturally aware, and globally capable language models. 🚀

We also evaluated our method on languages not seen during pre-training🌍: while performance is higher for seen languages, our transformations significantly improve both groups over the baseline—and in some cases are competitive with the teacher model📈(over 3x the student’s size).

📊 By inspecting the data itself, we see clear gains in quality along the targeted dimensions. Even when the interventions are relatively small, they produce substantial changes in completions improving their fluency, diversity, and difficulty ✨

⛰️With these simple transformations, we’re able to obtain consistent improvements across our 12 target languages and a diverse set of benchmarks, with particularly pronounced gains on open-ended tasks — our best proxies for real human use 💬

Only relying on translation often yields unnatural, Western-centric, and linguistically flat prompts.
💡We propose a simple, easy-to-implement solution to this problem:
🌐Transform translated prompts along three axes: Naturalization, Cultural Adaptation, and Difficulty.

🌍Most multilingual instruction data starts as English and translation can’t capture cultural nuance or linguistic richness
What if we optimized prompts instead of completions?
That’s the focus of our most recent work on prompt space optimization for multilingual synthetic data🗣️

🚀 Global MMLU Lite is now live on Kaggle Benchmarks!

Developed by @cohereforai.bsky.social, it spans 16 languages with both Culturally Sensitive & Agnostic samples - helping researchers uncover cultural & linguistic biases in multilingual evaluation.

Leaderboard: https://www.kaggle.com/benchmarks/cohere-labs/global-mmlu-lite/leaderboard

Notebook: https://www.kaggle.com/code/shivalikasingh95/global-mmlu-lite-sample-notebook