A team of researchers at MIT and other institutions has developed a new approach to guide large language models (LLMs) in generating accurate and error-free computer code. The method uses a probabilistic approach, boosting computational efficiency while improving the accuracy of LLM-generated outputs.
The current challenge with LLMs is that they often fail to follow the rules of programming languages or cause computers to crash. Existing methods for controlling the output of LLMs either distort their intended meaning or are too time-consuming to be feasible for complex tasks.
To address this issue, the researchers have engineered knowledge into LLMs to steer them toward the most promising outputs. This approach allows LLMs to allocate efforts toward outputs that are most likely to be valid and accurate, while discarding unpromising outputs early in the process.
The new architecture uses a technique called sequential Monte Carlo, which enables parallel generation from an LLM to compete with each other. The model dynamically allocates resources to different threads of parallel computation based on how promising their output appears.
In testing, the researchers’ method outperformed existing approaches, enabling small LLMs to generate accurate outputs for several real-world use cases, including molecular biology and robotics. This approach could also help non-experts control AI-generated content, such as businesspeople writing complex queries in SQL.
The development of this new architecture has significant implications beyond research, potentially improving programming assistants, AI-powered data analysis tools, and scientific discovery platforms. The researchers believe that their method can provide a more accurate and efficient way to generate code, enabling LLMs to make better decisions about the output they produce.
Source: https://news.mit.edu/2025/making-ai-generated-code-more-accurate-0418