Cognizant's AI Lab unveils new fine-tuning method for LLMs and reaches 61 U.S. patents

Cognizant's Artificial Intelligence (AI) Lab introduced a novel method for fine-tuning large language models (LLMs), aimed at reducing training costs and enhancing model efficiency. This advancement builds on the lab's ongoing research into AI, supported by the recent issuance of two additional U.S. patents, increasing its total to 61 patents.

The new research titled “Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning” highlights the application of evolution strategies (ES) for optimizing LLMs with billions of parameters. Traditional reinforcement learning (RL) methods often incur high costs and data requirements, while the ES approach promises enhanced performance and reduced resource utilization.

Incorporating this ES-based fine-tuning method, Cognizant's AI Lab aims to simplify hyperparameter tuning and improve the robustness of Large Language Model (LLM) training. The lab has already achieved a tenfold increase in processing speed since releasing the initial ES fine-tuning code, with plans to further scale the method for more complex tasks across larger models.

Alongside this research, the lab secured two new patents. The first patent focuses on AI-driven optimized decision-making systems for epidemiological modeling, while the second covers evolved data augmentation techniques that improve model performance with limited datasets. According to Risto Miikkulainen, VP of Research at Cognizant and a collaborator on the patents, both innovations facilitate effective model training despite challenges in dataset size.