Welcome to the 📖 Reading to Learn (R2L) Lab 🤖! We are in the Cheriton Department of Computer Science at the University of Waterloo.

The R2L Lab explores how language understanding improves machine learning efficiency and generalization. Our vision is to build intelligent, generalist agents that reason and act effectively across digital and physical environments. We focus on how language grounding enables efficient generalization, allowing systems to rapidly adapt to novel tasks.

Building and Benchmarking SOTA Agents: We develop agents and industry-standard benchmarks for complex workflows. Our platforms, including Spider 2 and OSWorld/Computer Agent Arena (computer control), serve as primary evaluations for OpenAI, Anthropic, Google, and Salesforce, driving progress in multimodal GUI grounding and operational knowledge.

Learning from Open Language Feedback: We use natural language to improve planning, replacing manual reward engineering with interpretable feedback and language-guided exploration. Our work on Language Feedback Models and Text2Reward demonstrates that language signals significantly enhance generalization and sample efficiency in robotics and continuous control tasks.

Synthetic Data & Efficient Learning: We advance model evaluation and training efficiency through synthetic data and curriculum learning. Key research threads include quantifying distributional differences between synthetic and real data distributions, developing sample-efficient automated curricula from expert teachers, and enabling scalable natural language querying over heterogeneous data lakes.