Large language models (LLMs) have long been celebrated for their prowess in textual reasoning, adeptly understanding and analyzing documents to provide logical answers. However, these same models often falter when tasked with solving even basic mathematical problems. This discrepancy arises because textual reasoning is not ideally suited for computational or algorithmic tasks. While some LLMs can generate code, such as Python, to tackle symbolic queries, they frequently struggle to determine when and how to use code effectively.
To address this challenge, researchers at MIT have developed CodeSteer, a smart assistant designed to guide LLMs in switching between text and code generation to provide accurate answers. CodeSteer, a smaller LLM itself, generates a series of prompts to iteratively steer a larger LLM, reviewing and refining its answers until the correct solution is achieved.
Enhancing LLM Accuracy with CodeSteer
CodeSteer’s introduction marks a significant advancement in improving the problem-solving capabilities of LLMs, particularly in tasks that are challenging to resolve through textual reasoning alone. The researchers found that augmenting a larger LLM with CodeSteer increased its accuracy on symbolic tasks, such as multiplying numbers, playing Sudoku, and stacking blocks, by more than 30 percent. This enhancement also enabled less sophisticated models to outperform more advanced ones with improved reasoning skills.
This development could have profound implications for complex tasks, such as generating paths for robots in uncertain environments or scheduling shipments in an international supply chain. According to Chuchu Fan, an associate professor of aeronautics and astronautics and principal investigator in the MIT Laboratory for Information and Decision Systems, “There is a race to develop better and better models capable of doing everything, but we’ve taken a complementary approach. We want to enable LLMs to select the right tools and methods, and make use of others’ expertise to enhance their own capabilities.”
A New Approach to LLM Training
Rather than retraining powerful LLMs like GPT-4 or Claude, MIT researchers opted to fine-tune a smaller, lightweight LLM to guide larger models between text and code. This approach avoids the risk of undermining the larger model’s other abilities. Yongchao Chen, a graduate student at MIT and co-author of the study, explains, “We were inspired by humans. In sports, a trainer may not be better than the star athlete, but the trainer can still give helpful suggestions. This steering method works for LLMs, too.”
CodeSteer functions as a trainer, reviewing queries to determine whether text or code is more suitable and generating prompts for the larger LLM to follow. If the initial answer is incorrect, CodeSteer continues to prompt the LLM to try different methods, such as incorporating search algorithms or constraints into its code, until the correct answer is found.
Breaking New Ground with SymBench
To fine-tune and test CodeSteer, researchers developed a new dataset called SymBench, comprising 37 complex symbolic tasks, including spatial reasoning, mathematics, order reasoning, and optimization. This dataset enabled CodeSteer to outperform all nine baseline methods evaluated, boosting average accuracy from 53.3 percent to 86.4 percent. Remarkably, the model maintained similar performance even on unseen tasks and across various LLMs.
Chen notes, “Our method uses an LLM’s own capabilities. By augmenting an LLM with the ability to smartly use coding, we can take a model that is already very strong and improve its performance even more.”
Future Directions and Expert Opinions
Looking ahead, the researchers aim to streamline CodeSteer to accelerate its iterative prompting process and explore how to effectively fine-tune a unified model capable of switching between textual reasoning and code generation. Jinsung Yoon, a staff research scientist at Google Cloud AI, who was not involved in the work, praises the research, stating, “This simple yet impactful method enables state-of-the-art LLMs to achieve significant performance improvements without requiring direct fine-tuning.”
Chi Wang, a senior staff scientist at Google DeepMind, adds, “Their success in training a smaller, specialized model to strategically guide larger, advanced models is particularly impactful. This intelligent collaboration among diverse AI ‘agents’ paves the way for more robust and versatile applications in complex real-world scenarios.”
This groundbreaking research is supported, in part, by the U.S. Office of Naval Research and the MIT-IBM Watson AI Lab. As the field of artificial intelligence continues to evolve, innovations like CodeSteer promise to enhance the application of LLMs across a diverse range of tasks, pushing the boundaries of what these models can achieve.
