Toward universal steering and monitoring of AI models

AI models can be "steered" by identifying and manipulating their internal concepts: Researchers developed a method to extract and adjust how AI models represent ideas internally, allowing them to expose vulnerabilities, reduce harmful outputs, and improve performance—similar to adjusting the "thinking" of a model rather than just changing its inputs.

Steering works better on newer, larger models: The study found that more advanced AI systems are actually easier to steer and control, which is encouraging for safety efforts, and steering can improve model capabilities beyond what standard prompting achieves.

Internal concept representations transfer across languages and can be combined: The same internal representations of concepts work across different human languages and can be combined to address multiple issues simultaneously, suggesting a universal underlying structure in how AI models store knowledge.

Monitoring harmful outputs is more effective using internal representations: Rather than judging whether model outputs are problematic after they're generated, predictive models built on the AI's internal concept representations can catch issues like hallucinations and toxic content more accurately—enabling earlier intervention.

Understanding AI's "internal knowledge" is key to both safety and capability: The research demonstrates that mapping how AI models internally represent concepts is foundational for both developing effective safeguards against misalignment and unlocking better model performance.