10. Out-of-Context Reasoning and Situational Awareness

This is the tenth lecture in the Language Models and Intelligent Agentic Systems course, run by Meridian Cambridge in collaboration with the Cambridge Centre for Data Driven Discovery (C2D3).

This lecture covers two emergent phenomena in language models, that may allow a language model to act differently depending on whether it is in training, evaluation, or deployment. Out-of-Context Reasoning is the ability to synthesise disparate pieces of information from pre-training or post-training data to reason about the world. This facilitates Situational Awareness, the ability for a language model to have self-knowledge, make inferences based on this knowledge, and act accordingly. Should we continue to be give models such detailed information about their deployment context?