Wenting Zhao: Reasoning in the Wild

In this talk, I will discuss my research on training and evaluating LMs to reason in real-world scenarios. First, I will present my research on better training strategies to teach LMs to reason. As user queries become more complex, the difficulty of collecting annotations grows rapidly. My work introduces training methods that enable LMs to learn reasoning without supervision. Results show that our approaches have great potential for improving model reasoning capabilities when human supervision is unavailable. Next, I will discuss my work on data collection in the wild. Here, I collect the WildChat dataset, which comprises 1 million naturally occurring conversations between users and chatbots. I will show that these "in-the-wild" conversations are more diverse and challenging than those in the academic benchmarks, providing a more accurate evaluation of LMs' capabilities when deployed in the real world. I will conclude the talk by outlining future research directions that could bring us closer to achieving superhuman reasoning capabilities in LMs.