Artificial intelligence (AI) researchers have discovered a method to guide the outputs of large language models (LLMs). By adjusting specific ideas or concepts hidden inside these models, the approach can make them more trustworthy and easier to train with less computer power. However, it also highlights weaknesses that could be exploited.
The work, led by UC San Diego and MIT, expands on earlier research from 2024 that introduced recursive feature machines, tools that spot mathematical patterns representing concepts within the models.
The researchers explain that simple math can change these patterns to influence the model's behavior. Experiments on open-source models like Llama and Deepseek targeted 512 concepts, such as fears, moods, and places. The technique worked across languages including English, Chinese, and Hindi, shedding light on the previously mysterious inner workings of these AI black boxes.
Improving performance while uncovering risks
This steering can enhance results for focused tasks, like converting code from one programming language to another, and spot hallucinations, which are inaccurate or invented facts produced by the AI. The researchers noted that working directly with the model's internal math offers advantages over just using natural language commands.
A downside is the possibility of jailbreaking, where the AI ignores safety rules and provides harmful advice, such as drug use instructions or false personal data. It can also amplify political biases or promote conspiracy theories, like claims about a flat Earth or dangerous vaccines.
The process is efficient, needing just one graphics processing unit, under a minute, and fewer than 500 examples. It suits larger open models but not closed ones like Claude. Future efforts aim to refine it for particular uses, potentially boosting both effectiveness and safety by better understanding the models' inner representations.
The researchers have described the methods and results of this study in a paper published in Science.