
@Cohere's research lab and open science initiative that seeks to solve complex machine learning problems. Join us in exploring the unknown, together.
Earth
Joined April 2019
- Tweets5.6K
- Following272
- Followers28.2K
- Likes3.7K
Pinned Tweet
We are committed to making meaningful progress in machine learning research through open collaboration. Follow this 🧵to stay on top of our research contributions.
New blog from our community! Catalyst Grant Recipient, @featuregraph (Nazia Habib), introduces us to FeatureGraph, a standardized approach to timeseries data analysis that creates a single source of truth for scientific and engineering research. 📈
Read the full post: labscommunity.cohere.com/blo…
Cohere Labs retweeted
My talk at @Cohere_Labs is now available on YouTube. Go watch it!
youtu.be/R75wEi3Ti-M?si=7x1c…
Interesting thing is the room was filled with MIT, Peen, OSU and other ivy leagues professors and phd students and then my poster (18y old high school senior) really grateful to get accepted and be a part of deepmath conference this year
Huge thanks to prof Zhihui Zhu and @OSUengineering for making this possible
Our community-led EdTech group is looking forward to hosting Waqas Aliemuddin for a presentation on "The Unfair Advantage of the Non-Technical Founder" when he will deep dive into the idea that founders without traditional software development backgrounds could possess an unexpected advantage!
Thanks to @Rafay_Mustafa1 for organizing this session 🔥
Join us on Thursday, November 19: luma.com/9rmpr89l
Cohere Labs retweeted
Extremely happy to host @savyakhosla for the audience to ask deep technical questions, discuss open problems, and explore future collaboration possibilities with the @Cohere_Labs research community
Scaling Visual Intelligence for High-Resolution and Long-Form Video
Abstract: What would it take for a robot or a wearable assistant to watch an entire day and help a blind person find their keys? Current vision systems make this prohibitively expensive because they split every image into a fixed grid of patches. As a result, doubling the resolution quadruples the token count, even when the added pixels carry no new information (e.g., a blank wall), and every video frame adds hundreds more tokens, even when it barely differs from the last. This talk presents region-based representations that encode each meaningful entity as a single token that persists across frames, so computation scales with what the scene contains rather than with resolution or video length. I will show how to generate region tokens efficiently, extend them to video, and align them with language. With region tokens, an hour of video takes fewer tokens than a patch-based encoder needs for a twenty-second clip, and performance improves on tasks ranging from segmentation and retrieval to episodic memory localization and spatial reasoning.
Bio: Savya Khosla is a Ph.D. student in Computer Science at the University of Illinois Urbana-Champaign, advised by Derek Hoiem and Alexander Schwing. His research focuses on efficient visual representations, long-video understanding, and multimodal learning. He develops region-based representations that enable vision models to perceive more detail and remember longer videos at a fraction of the compute cost. His work has resulted in multiple publications at CVPR, NeurIPS, ECCV, and ACL. Before his Ph.D., he was a software engineer at Google, working on Search, and has since conducted research at Meta, Adobe, Mila, NUS, and Ai2
Starting in an hour - 11 AM EDT / 08:00 PST / 21:30 IST
luma.com/17rxiswp?tk=UactKY
Join the @Cohere_Labs Discord for more - sites.google.com/cohere.com/…
Schools are embracing technology, but is it truly benefiting students? 🎒
Anshika Verma explores the vital difference between EdTech and PedTech. Discover how building tech around real learning outcomes can change education: youtu.be/_uE_8ORjIvA
Cohere Labs retweeted
Excited to speak at the @Cohere_Labs Open Science Community next week on "Kernel Autoresearch for Open-Ended Model Discovery" 🧬
Looking forward to sharing our work and discussing agentic & open-ended approaches to automated model discovery.
Hope to see you there!
🔗 luma.com/8po3irq5
Our Open Science Community is excited to share our events for next week, November 9 -14! Check out the upcoming sessions:
📆 Kernel Autoresearch for Open-Ended Model Discovery with @richardcsuwandi: luma.com/8po3irq5
📆 Context Language Models with @RulinShao: luma.com/f0v9tsvj
Thanks to our Open Science community leads @Sree_Harsha_N, @KanwalMehreen2 and @AhmadMustafaAn1 for organizing these events!
Our Open Science Community is excited to share our events for next week, November 9 -14! Check out the upcoming sessions:
📆 Kernel Autoresearch for Open-Ended Model Discovery with @richardcsuwandi: luma.com/8po3irq5
📆 Context Language Models with @RulinShao: luma.com/f0v9tsvj
Thanks to our Open Science community leads @Sree_Harsha_N, @KanwalMehreen2 and @AhmadMustafaAn1 for organizing these events!
Tomorrow we are hosting a research recap for our CulturalRiddles project.
Hope you can tune in to learn about the genesis of this work, collaboration across 150+ people, research findings, and potential future directions.
luma.com/k0uqq8iq
Tune in on November 5th to learn about the CulturalRiddles project and how this benchmark of 6,200 community-written riddles across 51 languages, spanning 45 countries was created to evaluate whether AI models can reason about the cultural context behind a question! 🧩
Join some of our project leads: @mychiffonn, @Ipouyall, Tianyi Hu, Henry Gagnier and @aroraakhilcs as well as, Julia Kreutzer (Staff Research Scientist, Cohere Labs) for a release recap covering the story behind the project, why it matters, and the key findings from our research.
Hear directly from the people who brought this global, community-driven effort to life!
📅 RSVP to join: luma.com/k0uqq8iq
Learn more about the CulturalRiddles benchmark: cohere.com/research/papers/c…
Curious about transforming AI models efficiently? ⚡
Anupam Nayak explores how model architecture can be changed late in training rather than remaining fixed after pretraining, catch the full talk: youtu.be/CtfjZ75Kyok
1
8
1,430
How can multi-agent systems boost efficiency while cutting costs? 🤔
@keenansamway, Research Scholar at Cohere Labs explores a novel approach to team construction in agent, resulting in a 70% cost reduction.
Watch the full session: youtu.be/kkTKk2cZ3y0
Learn more in the blog post: cohere.com/blog/lessons-from…
Celebrating our October Cohere Labs Community Lead anniversaries! 🎉
@katomubirusteve launched our Africa regional program and continues to make an impact both inside and outside the lab, including through his participation in Expedition Aya events.
@cole__ai helped launch our Embodied AI program and has welcomed numerous speakers to share their work with our community.
We’re grateful to both Cole and Stephen for their leadership, initiative, and continued contributions to our global open science community. 💙
Tune in on November 5th to learn about the CulturalRiddles project and how this benchmark of 6,200 community-written riddles across 51 languages, spanning 45 countries was created to evaluate whether AI models can reason about the cultural context behind a question! 🧩
Join some of our project leads: @mychiffonn, @Ipouyall, Tianyi Hu, Henry Gagnier and @aroraakhilcs as well as, Julia Kreutzer (Staff Research Scientist, Cohere Labs) for a release recap covering the story behind the project, why it matters, and the key findings from our research.
Hear directly from the people who brought this global, community-driven effort to life!
📅 RSVP to join: luma.com/k0uqq8iq
Learn more about the CulturalRiddles benchmark: cohere.com/research/papers/c…
Cohere Labs retweeted
I will be giving a talk at @Cohere_Labs on our latest research from my internship at @Apple .
Title: "𝗩𝗶𝘀𝘂𝗮𝗹 𝗚𝗿𝗼𝘂𝗻𝗱𝗶𝗻𝗴 𝗦𝗮𝗳𝗲𝘁𝘆 𝗶𝗻 𝗩𝗶𝘀𝗶𝗼𝗻-𝗟𝗮𝗻𝗴𝘂𝗮𝗴𝗲 𝗠𝗼𝗱𝗲𝗹𝘀"
Luma event: luma.com/rodoya2r?tk=LwFJBG
Our Open Science Community is excited to share our events for next week, November 2 -7! Check out the upcoming sessions:
📆 When the Update Is Not a Gradient: Input-to-State Stability of Self-Repairing Neural Network Training with Rokib Al Dhin Raadh: luma.com/p28sna3q
📆 Harness²: Recursive Agent Harnessing for an Open World with @ruiyao_xu: luma.com/wpodv4w1
📆 Visual Grounding Safety in Vision-Language Models with @Erf_Shayegani: luma.com/rodoya2r
Thanks to our Open Science community leads @KatrinaL2899, @KanwalMehreen2, @AhmadMustafaAn1 Damani, and Manuel Villanueva for organizing these events!
Don't forget to tune in tomorrow to hear from @keenansamway, @mrdanieldsouza and @nouhadziri🔥
RSVP to join: luma.com/i443qd6c
📜 Read the paper: cohere.link/tStRYJS
✍ Check out the blog post: cohere.link/0TzgKoy
Join @keenansamway, @mrdanieldsouza, and @nouhadziri as they deep dive into their latest work "Scaling Inference-Time Compute with Teams of Coding Agents" and they'll discuss:
🧠 Strategic delegation and agent team design
💸 Achieving near-optimal performance at 70% lower cost
⚡ Balancing compute, cost, and latency
Register to join the session on Friday, October 30th: luma.com/i443qd6c
As part of our commitment to open science, the CulturalRiddles data is available on @huggingface so researchers and the community can explore how AI models understand culture through riddles. 🤗
Check out the dataset, dig into the data, and build on this work.
We’re excited to see what you discover: huggingface.co/datasets/Cohe…
Our Open Science Community is excited to share our events for next week, November 2 -7! Check out the upcoming sessions:
📆 When the Update Is Not a Gradient: Input-to-State Stability of Self-Repairing Neural Network Training with Rokib Al Dhin Raadh: luma.com/p28sna3q
📆 Harness²: Recursive Agent Harnessing for an Open World with @ruiyao_xu: luma.com/wpodv4w1
📆 Visual Grounding Safety in Vision-Language Models with @Erf_Shayegani: luma.com/rodoya2r
Thanks to our Open Science community leads @KatrinaL2899, @KanwalMehreen2, @AhmadMustafaAn1 Damani, and Manuel Villanueva for organizing these events!
We're thrilled to have some of our latest work featured at the ORACLE Workshop at EMNLP 2026 today!
Poster Presentations:
🌐 The Culture Funnel - cohere.com/research/papers/t…
🌉 Building Multilingual Bridges - cohere.com/research/papers/b…
Oral Presentation:
🏆 The IOL-AI Challenge - cohere.com/research/papers/t…
Congrats to authors AnanyaSahu , @mhrnz_m, @mrdanieldsouza, @rsrchvslstrytlr, Julia Kreutzer, @mziizm, @alexrs95, Alexandre Bérard, Eduardo Sánchez, Rita Berrada, @DanMirea4, Sara Rajaee, Alexander Piperski, Ana Meta Dolinar, Boris Iomdin, Andrey Nikulin, Mariya Shmatova👏
Proud to see our @cohere colleagues work highlighted at the Impact Speech Conference at @emnlpmeeting! 🤩
Explore all of our research: cohere.com/research/papers
4
749
Cohere Labs at the MRL Workshop @emnlpmeeting: CulturalRiddles: A Multicultural Benchmark of Riddles
The CulturalRiddles benchmark tests AI models on 6,200 culturally grounded riddles across 51 languages from 45 countries, revealing that models struggle significantly with cultures outside their training data, hallucinate rather than admit uncertainty, and fail even with translated content. This open collaboration project highlights that cultural knowledge—distinct from written facts—remains a frontier current AI models have yet to master.
📜 Read the paper: cohere.com/research/papers/c…
Congratulations to all authors, collaborators and riddle writers on this project!
The @cohere and Cohere Labs team are proud to be a part of the WMT conference at @emnlpmeeting! 👏
Explore all of our research: cohere.com/research/papers
18
1,236



















