NLP & Agents

Language models are no longer just chatbots. With the right harness, meaning the tools, memory, workflows and safeguards around the model, they can carry out long, multi-step tasks on their own. That harness often matters as much as the model itself, and it is where we focus our work.

We work in small teams of students who pick projects they find genuinely exciting. Some teams pursue research questions with the aim of publishing their results. Others build tools and systems that are useful to ReLU and to students. Members come from different backgrounds and experience levels, and every team mixes newcomers with more experienced members, so everyone learns along the way.

We take engineering seriously, even for internal projects. Our teams use version control, code review, automated testing and systematic evaluation of AI output, the same practices used in industry. Because agents can be unpredictable, we put particular emphasis on testing how reliable they are and running them safely in sandboxed environments.

The lab runs open models on our own GPU hardware, which lets us experiment freely, run many agents in parallel, and keep data under our own control. Alongside the project work, we host hackathons and competitions where members build agents that compete against each other. It’s a fun way to learn, and a good excuse to get together.