Cognizant pioneers evolution strategies to fine-tune billion-parameter LLMs, surpassing reinforcement learning.
Cognizant’s AI Lab announced a novel approach to fine-tuning large language models (LLMs) using evolution strategies (ES), marking a significant advancement beyond traditional reinforcement learning (RL). This method applies a population-based search to optimize data augmentation and selection automatically, enabling more scalable, stable, and cost-effective training of models with billions of parameters. The team demonstrated, for the first time, error-free performance on million-step reasoning tasks, overcoming previous limitations where errors compounded in long dependency chains. These innovations include a 10x speed-up in evolutionary optimization using modern vision-language model inference engines, illustrating substantial improvements in efficiency.
This research reflects a transformative shift in LLM fine-tuning paradigms, with the newly issued US patent covering these evolved data augmentation techniques. The developments were achieved by a multidisciplinary team led by Drs. Meyerson, Liang, Qiu, and Professor Miikkulainen. They reinforce Cognizant’s leadership in AI research by pushing the boundaries of large-scale, high-accuracy, and computationally efficient model training. These methods promise broader applications in complex generative AI tasks, offering a foundation for next-generation autonomous agents and decision-making systems.
Source: Cognizant AI Lab