SelfIE: Self-Interpretation of Large Language Model Embeddings
This work addresses the need for reliability and transparency in LLMs by providing tools to explain and control their reasoning processes, which is crucial for developers and users in AI safety and interpretability.
The paper tackles the problem of interpreting and controlling large language models (LLMs) by proposing SelfIE, a framework that enables LLMs to interpret their own embeddings in natural language, revealing internal reasoning in cases like ethical decisions and harmful knowledge recall, and it introduces methods like Supervised Control and Reinforcement Control for editing concepts and erasing harmful knowledge without supervision targets.
How do large language models (LLMs) obtain their answers? The ability to explain and control an LLM's reasoning process is key for reliability, transparency, and future model developments. We propose SelfIE (Self-Interpretation of Embeddings), a framework that enables LLMs to interpret their own embeddings in natural language by leveraging their ability to respond to inquiries about a given passage. Capable of interpreting open-world concepts in the hidden embeddings, SelfIE reveals LLM internal reasoning in cases such as making ethical decisions, internalizing prompt injection, and recalling harmful knowledge. SelfIE's text descriptions on hidden embeddings also open up new avenues to control LLM reasoning. We propose Supervised Control, which allows editing open-ended concepts while only requiring gradient computation of individual layer. We extend RLHF to hidden embeddings and propose Reinforcement Control that erases harmful knowledge in LLM without supervision targets.