As Large Language Models (LLMs) are used increasingly in research, creativity, and decision-making, it becomes crucial to ensure that their outputs are safe, reliable, and accurate. Yet, they often misbehave: they may memorize text, hallucinate responses, be jailbroken, or exhibit flawed logical reasoning.
Previous work has shown that the internal computational traces of LLMs, namely their hidden states, carry useful signals about different types of LLM misbehavior, e.g., the truthfulness of model outputs, and can be exploited to detect such misbehaviors.
However, these existing approaches largely ignore the structure and complexity of these traces and therefore fail to incorporate appropriate inductive biases for optimal performance. This talk presents a line of work on characterizing data structures in LLM computational traces and developing tailored architectures that learn from them, including final-layer outputs, intermediate activations, and attention matrices.
Our approaches identify misbehaviors more effectively than prior methods and generalize across tasks and LLMs, allowing techniques learned on one task or model to transfer to others.