Using AutoResearch for Science

​Abhinav Kumar is a Researcher at Lossfunk, where he has been extending Karpathy's auto-research framework to test whether neuroscience priors can improve performance on tasks like BabyLM. He has previously worked on building an AI Research Mentor and probing sycophancy vectors in LLMs.

​In this session, he will walk through Karpathy's auto-research, a setup where an LLM agent autonomously iterates on a problem against a editable markdown file running for hours (or weeks) on a single task.

​We will look at how the community has been using auto-research across CS/ML for program discovery, fine-tuning, architecture discovery and more use-cases across domains. We will also walk through Lossfunk's experiments, and discuss the failure modes along with best practices for using the system.

​The session will end with a discussion of CAISc, the verifiable problems which attendees can attempt on their own, as well as recommendations for open-ended research problems suitable for using auto-research.

​Karpathy's Auto-Research repo - https://github.com/karpathy/autoresearch

​CAISc Verifiable Problems - https://caisc2026.github.io/verifiable-problems/

​Part of the CAISc 2026 pre-conference series. Learn more at caisc2026.github.io.