New AI Steering Method Controls Internal Model Concepts

Researchers from UC San Diego and MIT have developed a mathematical method to steer large language model (LLM) outputs by modifying internal concept patterns ^[1]. Published in the journal Science on February 19, 2026, the technique allows developers to influence model behavior directly through internal representations ^[2]. This breakthrough could lead to more reliable and adaptable artificial intelligence systems ^[1]. This approach relies on predictive algorithms to identify specific semantic patterns, such as mood or geographic location, within the model's layers ^[2][3]. By adjusting these mathematical vectors, researchers can improve task performance—like code translation—without the high cost of retraining the entire network ^[1][3]. The method requires significantly less computational power than existing training techniques ^[2]. While the method enhances reliability and reduces hallucinations, it also exposes system vulnerabilities ^[2]. The team successfully steered models to provide restricted information, such as drug manufacturing instructions, highlighting the need for robust AI safety frameworks ^[1][2]. Researchers have made their code public to encourage further safety exploration ^[3].