Access to the resources needed to conduct machine learning research, such as compute and state-of-art large language models (LLMs), is not always easily available for researchers. Not all research institutions have their own compute clusters or expertise to run models, and many existing research grants don’t allow researchers to spend their funds on accessing proprietary models. This is especially acute for researchers outside of well-funded institutions and labs (largely in North America), but it is also the case for many researchers outside of computer science departments who want to apply LLMs to their research fields – from health to education.

In an effort to help narrow this gap, Cohere For AI launched our Research Grant program in July 2023. These grants provide academic researchers and developers with subsidized access to the Cohere API to support their research into advancing safe, responsible LLM capabilities and applications.

Since then, we have awarded 75+ grants to researchers across 18 countries, such as Canada, France, Argentina, Finland, and Indonesia. Our grantees come from top global institutions, including the University of Oxford, MIT, and Copenhagen University, as well as independent research groups and international organizations.

We've awarded grants to projects that span a range of research areas, including:

Language model safety, including bias, explainability, hallucinations, toxicity, and adversarial testing
Language model applications for prosocial goals in fields such as education, climate science, content moderation, healthcare, law, history, and social science
Language model capabilities, such as improving information retrieval accuracy or model efficiency
Multilingual capabilities, such as increasing language model performance and safety in languages beyond English
Values alignment exploring how to ensure generative AI model behavior meets people’s expectations, preferences and values

Three Research Projects by Our Grantees

Building LLM tools for teachers
Jussi Jauhiainen and Agustín Garagorry, University of Turku, Finland

*A diagram of a model developed by Jussi, Agustín, and their colleagues*

Teachers and learners around the world are experimenting with AI tools – particularly language models – to explore new ways to approach education. Jussi, Agustín and their colleagues recognised that much of this experimentation is limited by a lack of good evidence regarding how to use these tools well and responsibly.

To develop more robust language model tools for education, Jussi and Agustín are using language models to adapt learning materials to match specific learners' levels, tailoring content to provide an incremental learning experience. Recently, they have studied multiple LLM comparisons for evaluating student performance, highlighting the need to choose the right tool, raise awareness of potential risks, and enhance informed decision-making by educational stakeholders. Currently working in the Finnish language, they plan to expand this approach to English, Spanish, and several African languages, in collaboration with multiple partners.

The Cohere For AI research grant allows Jussi, Agustín, and their colleagues to run their experiments on Cohere’s models, expanding their test cases and offering them a broader evidence base for their work.

Read some of their previous research:

Aligning language model behavior with human values
Hannah Rose Kirk and colleagues, University of Oxford, UK

Human feedback is an essential component in the development and improvement of LLMs. But there are many open questions about whose feedback is included – and whose values language models are aligned with.

To address these questions, Hannah and her colleagues developed PRISM, a dataset that captures the preferences and feedback of 1,500 individuals from 75 countries in conversation with 21 LLMs.

In this work, the team included a wide geographic and demographic range, ensuring a diverse set of perspectives and preferences. In addition the team linked individual feedback to detailed participant profiles, allowing for personalized insights and the identification of patterns based on user characteristics.

Through this work, Hannah and her colleagues aim to enhance how LLMs can be better aligned with users’ needs and expectations, ensuring that these models are not only accurate but also responsive to the diverse global audience they serve.

The Cohere For AI research grant allowed the team to include Cohere’s Command models in their experiments. The models not only performed well, but also offered key insights, such as how users' preferences often relate to the length and tone of a model’s output.

Find out more about this project:

Understanding stereotypes in LLMs in Argentina
Luciana Benotti and Guido Ivetta, Universidad Nacional de Córdoba, Argentina

*A photo from the team’s workshop on 1 June 2024*

Assessing biases in AI models – such as when language models reflect harmful stereotypes or discriminate unfairly between groups of people – often excludes discrimination experts due to technical barriers. To address this, Luciana, Guido, and their colleagues have developed a tool and designed a methodology that allows social science communities and domain experts in Latin America to explore biases and discriminatory stereotypes present in word embeddings and language models.

Now, the team is building an ecosystem of secondary school teachers and their students, with training in bullying and harassment management, to gather community-built datasets that represent stereotypes in their country, Argentina. These datasets will be the keystone to detecting and characterizing discriminatory behaviors and hate speech in language model outputs used in Argentinian settings. This is crucial, as many existing approaches towards understanding bias in language models are oriented towards English language contexts.

The Cohere For AI research grant will allow Luciana and Guido to include Cohere’s models alongside other open source models in the tool that their community will use to explore bias, and to test Cohere’s models against a benchmark they are developing to assess how stereotype biases are present in language models.

Three Research Projects by Our Grantees

Building LLM tools for teachersJussi Jauhiainen and Agustín Garagorry, University of Turku, Finland

Aligning language model behavior with human valuesHannah Rose Kirk and colleagues, University of Oxford, UK

Understanding stereotypes in LLMs in ArgentinaLuciana Benotti and Guido Ivetta, Universidad Nacional de Córdoba, Argentina

Building LLM tools for teachers
Jussi Jauhiainen and Agustín Garagorry, University of Turku, Finland

Aligning language model behavior with human values
Hannah Rose Kirk and colleagues, University of Oxford, UK

Understanding stereotypes in LLMs in Argentina
Luciana Benotti and Guido Ivetta, Universidad Nacional de Córdoba, Argentina

Granting Access: Supporting Researchers to Use LLMs