Research Idea Submission
Hello Tomek Korbak! I applied to MATS program at the very last second - please excuse the late submission!
Below are my responses.
What do you want to work on? Please give a 1-2 paragraph pitch for your research idea that fits this stream.
I want to investigate LLM latent reasoning, focusing specifically on whether this internal capability can be controllably harnessed and enhanced using new methods. Critically, understanding how to enhance latent reasoning provides invaluable insight into its underlying mechanisms, scalability, and ultimately, its controllability-essential knowledge for safety, risk assessment, and developing robust control measures for future systems. Therefore, I propose employing controlled enhancement not just for capability exploration, but primarily as an experimental probe to empirically study these factors, generating publishable findings on the characteristics and manageability of non-explicit reasoning to better evaluate and guide the development of safer advanced LLM agents.