At Cohere Labs, we build open research pathways that don’t assume proximity, pedigree, or privilege. Many researchers discover our open science community at a time of transition. They’re seeking mentorship, support, and, often, a clearer sense of how to begin. They arrive with curiosity and ambition, but without a roadmap.

This story is about what can happen when those early moments are met with opportunity, trust, and space to grow.

Ram Kadiyala and Jebish Purbey were once those researchers, entering Cohere Labs’ open science community with ambition, curiosity, and more questions than answers.

Just two years later, they’re creating the very pathways they once sought, opening doors for others, leading global collaborations, and reshaping inclusive, action-oriented AI research.

Finding Opportunity Beyond Borders

Jebish joined our open science research community from Nepal, where opportunities to participate in cutting-edge ML research were limited locally. What he found was a global network: researchers working across borders, languages, and disciplines, united by a shared belief in open, collaborative science and a desire to change where breakthroughs happen.

His journey began with contributions to the Cohere Labs INCLUDE and Kaleidoscope projects. There, Jebish collaborated on the construction of evaluation suites and co-authored papers focused on building robust benchmarks for multilingual large language models. These benchmarks aimed to reflect real-world language use and span diverse languages, rather than narrow, English-centric evaluations. These works spanned borders and languages, but were united by a shared belief that good research should reflect the world it serves.

Meeting Through the Work

Some collaborations begin with introductions; others begin with shared effort. Ram arrived to the Cohere Labs Open Science Community from a different angle. He was actively looking for collaborators and meaningful projects, and he stepped into a leadership role as a project captain in Expedition Aya, a global challenge designed to foster collaboration and push the boundaries of multilingual AI research. His team’s project focused on multilingual AI-generated text detection, an increasingly urgent problem as generative systems scale globally.

During Expedition 2024, Ram and Jebish worked on separate projects, but they observed something familiar in each other’s work: clarity of purpose, care for impact, and a bias toward doing rather than over-planning.

This observation would later become a core lesson Jebish shares with newer researchers:

Admiration led to outreach. Outreach led to a shared vision. And from that, a new collaboration was born.

Building the ML Agents Group

The ML Agents Group formed around a simple idea: research moves faster and farther when people are invited in. Together, Ram and Jebish founded the group, grounded in four shared values:

Collaborate broadly: invite people in, rather than gatekeeping.
Address real-world problems: focus on work that matters outside a research environment.
Align with diverse communities across geographies, cultures, and backgrounds.
Be action-oriented: prioritize building, testing, and feedback over endless literature reviews.

The final value is perhaps the most distinctive. Instead of waiting for perfect framing, the group learned by building.

Proof in the Results

Over time, this way of working began to compound. In a short span of time, the group has built a body of research that spans multilingual safety, cultural representation, and benchmarking, much of it emerging from community projects and evolving through continued collaboration.

During the 2024 Expedition, the group developed Robust and Fine-Grained Detection of AI-Generated Texts, a multilingual study that goes beyond binary classification to explore more realistic detection settings. While still seeking a home, Ram and Jebish consider this their strongest technical contribution from this time, shaping much of the group’s thinking on robustness and evaluation.
Momentum continued in the 2025 Expedition with Uncovering Cultural Representation Disparities in Vision-Language Models, a collaboration that examines how vision-language systems represent culture across contexts. The paper was accepted to AACL 2025 Findings.
Between the 2024 and 2025 Expeditions, the group produced Improving Multilingual Capabilities with Cultural and Local Knowledge in Large Language Models While Enhancing Native Performance, addressing a key challenge in multilingual modeling. This work was accepted to AACL 2025 Main Conference (Oral).
Following the 2025 Expedition, the ML Agents Group released its first collective paper, DSBC: Data Science Task Benchmarking with Context Engineering, which introduces a new approach to benchmarking data science tasks. The paper was accepted to AACL 2025 Main.
The group is now developing its next project, Do All Tasks Benefit Equally from Reasoning?, an idea sparked during the 2025 Expedition and now being shaped into a full paper.

Together, these projects demonstrate a research culture rooted in experimentation, collaboration, and shared ownership, where ideas are seeded in community spaces, refined through action, and carried forward by growing teams of contributors.

Growing the Next Leaders

As the group matured, so did its sense of responsibility. Leadership became less about directing and more about creating room for others to step forward. Heading into 2026, Ram and Jebish are evolving how they collaborate. They are actively welcoming collaborators into leadership roles, continuing the cycle of mentorship that supported them early on.

Jebish recently welcomed Ankan Ghosh and Sovit Rath to the ML Agents leadership team. Together, they will spearhead the exploration of novel research avenues over the coming months, with a particular focus on integrating agents into open maps. The team remains open to new collaborators and individuals eager to deepen their understanding of research on and utilizing agentic ML.

Meanwhile, Ram is partnering with Suman Debnath to launch new multilingual research initiatives within the community, including an open-source LLM project focused on South Asian languages, with GPU support from Anyscale. These projects are just about to get started, making now an ideal moment to join.

The Cycle Continues

In just two years, Ram and Jebish have moved from newcomers with limited experience to leaders shaping the future of collaborative AI research. They received support, stepped into responsibility, and now uplift others in turn.

This is only one story, but it reflects a larger truth we see again and again: the future of AI research is collaborative, global, and community-driven.

If you’re hoping to make 2026 the year you step into your path in ML research, let this be your inspiration and your invitation.

How the Cohere Labs open research community turns early-career researchers into global leaders