A McGill University-led team of researchers has developed a new method that significantly improves the accuracy and efficiency of large language models (LLMs) when generating computer code and other structured text.
Large language models (LLMs) are powerful AI tools that are capable of writing everything from essays to programming scripts in a wide range of languages. However, they often make mistakes when generating text that must conform to constraints, leading to errors or outputs that don’t work. Existing solutions to fix these issues are often unreliable or too slow.
The new approach was developed by a team led by McGill University’s Timothy J. O’Donnell that includes researchers from MIT, ETH Zürich, and Yale. It uses a technique called Sequential Monte Carlo to co-ordinate multiple outputs and prioritize the best ones at each step, discarding unpromising outputs early in the process.
This boosts computational efficiency, saving time, reducing errors and allowing even smaller AI models to perform better than much larger ones. Their method allows an LLM to allocate efforts toward outputs that are most likely to be valid and accurate.
“As a linguist, one of the things that is exciting is that this work provides a framework where we can start to formulate real theories of meaning. We are going beyond LLM models for words, to symbolic models of their underlying meaning said O’Donnell, who is an associate professor and William Dawson Scholar in the Department of Linguistics at McGill and a core academic member of Mila-Québec Artificial Intelligence Institute, where he holds a Canada CIFAR AI Chair.
This breakthrough has potential applications across programming assistance, data analysis, robotics and scientific research, and represents a step toward making AI more understandable and controllable for non-experts.
The members of the research group (“GenLM Consortium”) say they plan to continue their work and will be releasing the software as an open-source toolkit in the near future.
The paper, “Syntactic and Semantic Control of Large Language Models via Sequential Monte Carlo,” was presented at the International Conference on Learning Representations (ICLR), April 24-28.