Cognizant's AI Lab Announces Breakthrough Research for Fine-Tuning LLMs and Records its 61st U.S. Patent Issuance

Rhea-AI Summary

Cognizant (Nasdaq: CTSH) announced that its AI Lab developed an evolution strategies (ES) method to fine-tune LLMs with billions of parameters, claiming improved scalability, stability and lower training-data needs versus reinforcement learning. The Lab reported a 10X speed-up after code and infrastructure refactors using faster vLLM inference. It also recorded two new U.S. patents—No. 12,406,188 (issued Sep 2, 2025) and No. 12,424,335 (issued Sep 23, 2025)—bringing the Lab's U.S. patent total to 61.

Positive

• 10X speed-up in ES fine-tuning codebase
• Two U.S. patents issued on Sep 2, 2025 and Sep 23, 2025
• Reached 61 U.S. patents in AI innovations
• Demonstrated ES fine-tuning on LLMs with billions of parameters

Negative

• None.

Insights

AI Research Strategist

New ES-based fine-tuning and two U.S. patents strengthen Cognizant's AI capabilities and may reduce LLM training costs.

Cognizant's AI Lab reports a novel fine-tuning method using evolution strategies (ES) that it says scales to multi‑billion‑parameter LLMs, claims improved stability versus reinforcement learning, and reports a 10X speed-up after infrastructure refactoring. The announcement pairs this research with two newly issued U.S. patents, bringing the lab's U.S. total to 61, dated September 23, 2025 and September 2, 2025.

The business mechanism is clear: a gradient‑free ES framework aims to reduce required training data and lower compute cost for post‑training alignment of LLMs, while the patents target epidemiological decision models and evolved data augmentation to improve model robustness with limited data. Key dependencies and risks include real‑world reproducibility, peer validation of comparative performance versus state‑of‑the‑art RL, and successful scaling to the largest models the Lab plans to target; those factors determine practical cost and quality gains.

Watch for formal benchmarks, open‑source code releases or peer‑reviewed results, and evidence of scaling to the largest models over the next 6–18 months; also monitor any product integrations or commercial announcements that show this research moving into customer solutions.

TEANECK, N.J., Oct. 28, 2025 /PRNewswire/ -- Cognizant (Nasdaq: CTSH) today announced a breakthrough from its AI Lab that introduces a novel, efficiency-focused method for fine-tuning large language models (LLMs) -- showing significant promise of reducing training costs compared to traditional approaches. At the same time, the lab continues to be recognized for its AI innovations with two newly issued U.S. patents that advance AI decision making and data augmentation.

"We're excited to be continuing to break new ground when it comes to AI innovation and be recognized for it by the US patent office," said Babak Hodjat, Chief AI Officer at Cognizant's AI Lab. "Our latest research, using evolution strategies (ES) for fine-tuning LLMs, has the potential to disrupt the industry. Our approach not only uses less training data than reinforcement learning (RL); it also makes the process more accurate, increasing the quality of work the AI can produce. It's an exciting time for our Lab."

In new groundbreaking research titled "Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning" (Xin Qiu, Yulu Gan, Conor F. Hayes, Qiyao Liang, Elliot Meyerson, Babak Hodjat, Risto Miikkulainen), Cognizant's AI Lab introduces the first successful use of evolution strategies (ES) to fine-tune LLMs with billions of parameters, marking a transformative new approach beyond traditional reinforcement learning (RL) methods. The net outcome has the potential to increase the effectiveness of training LLMs for a given task while significantly reducing the required training data and associated cost.

Every major LLM must be fine-tuned to become more aligned and useful for the task it is being asked to address, similar to how a student eventually graduates in a particular specialist domain. Reinforcement Learning (RL) is the current preferred method for training LLMs but it is expensive, hard to scale and sometimes the AI learns to "game the system" instead of producing high-quality work. Using evolution strategies – a gradient-free optimization algorithm that directly searches in parameter space – Cognizant's AI Lab was able to overcome the major limitations of RL-based fine-tuning methods. This innovative framework demonstrated improved performance compared to state-of-the-art RL techniques, delivering greater scalability, efficiency, and stability. As LLMs grow increasingly complex, this ES-based fine-tuning approach marks a major leap forward toward more reliable, adaptable, and efficient post-training by simplifying hyperparameter tuning and improving robustness. Since our initial ES fine-tuning code release, we have optimized our codebase and achieved a 10X speed-up, by refactoring the infrastructure with faster vLLM inference engines. The Lab's next step is to scale its method to fine-tune the largest available LLMs across a range of complex tasks.

In addition to this breakthrough, the lab cemented other important innovations, with two new U.S. patents being granted, bringing its U.S. total to 61. The two latest patents highlight key innovations from Cognizant's AI research:

• U.S. Patent No. 12,424,335 (issued September 23, 2025) covers systems and methods for AI-based optimized decision making for epidemiological modeling. It describes the use of neural networks to predict epidemiological trends, including applications such as COVID-19, by combining separate LSTM models for case and intervention histories into a unified predictor. By enforcing real-world constraints, the approach aims to improve forecast accuracy even with limited data, advancing deep learning applications in public health.
• U.S. Patent No. 12,406,188 (issued on September 2, 2025) describes a systems and methods for evolved data augmentation and selection, utilizing population-based search to automatically discover and select optimal data augmentation operations. This approach is designed to improve model robustness and performance, even with limited datasets, enhancing the efficiency and reliability of AI across real-world applications.

These innovations, developed by Cognizant's researchers including Dr. Jason Liang, Dr. Elliot Meyerson, Olivier Francon, Dr. Xin Qiu and Professor Risto Miikkulainen, reinforce Cognizant's leadership in pushing the boundaries of AI and machine learning.

"While deep learning can be applied to many real-world domains, its transformative power often comes out only when there are millions of data points to train the models," said Risto Miikkulainen, VP of Research and Professor of Computer Science at UT Austin. "Obtaining such large datasets is often unrealistic. These two innovations enable effective model training with fewer examples, expanding the applicability of deep learning."

About Cognizant
Cognizant (Nasdaq: CTSH) engineers modern businesses. We help our clients modernize technology, reimagine processes and transform experiences so they can stay ahead in our fast-changing world. Together, we're improving everyday life. See how at www.cognizant.com or @cognizant.

About the Cognizant AI Lab
The mission of the Cognizant AI Lab is to maximize human potential with Decision AI, a form of AI that combines generative AI, multi-agent architecture, deep learning, and evolutionary AI to create sophisticated decision-making systems. Decision AI powers Cognizant's Neuro® AI platform, which is utilized by Fortune 500 companies and non-profits to discover new ways to exceed their goals. The platform enables organizations to rapidly build AI that optimizes decision-making, leading to revenue growth and societal progress.

Led by AI pioneers Babak Hodjat and Risto Miikkulainen, the lab collaborates with institutions, academia, and technology partners to develop groundbreaking AI solutions responsibly. With over 120 patent filings globally (including issued and pending applications), the lab excels at combining scientific innovation with commercial application. It supports Cognizant's goal of improving everyday life, focusing on business and AI-for-good applications.

U.S. Name: Gabrielle Gugliocciello Email: gabrielle.gugliocciello@cognizant.com

Europe / APAC Name: Sarah Douglas Email: sarah.douglas@cognizant.com

India Name: Vipin Nair  Email: Vipin.Nair@cognizant.com

 

View original content to download multimedia:https://www.prnewswire.com/news-releases/cognizants-ai-lab-announces-breakthrough-research-for-fine-tuning-llms-and-records-its-61st-us-patent-issuance-302597082.html

SOURCE Cognizant

FAQ

What did Cognizant (CTSH) announce on October 28, 2025 about LLM fine-tuning?

Cognizant announced an ES-based fine-tuning method for LLMs with billions of parameters that the Lab says improves scalability and reduces training-data needs versus RL.

How much faster is Cognizant's ES fine-tuning after optimization?

The Lab reported a 10X speed-up in its ES fine-tuning codebase after refactoring and using faster vLLM inference.

Which U.S. patents did Cognizant receive related to AI in 2025?

Cognizant received U.S. Patent No. 12,406,188 (issued Sep 2, 2025) and No. 12,424,335 (issued Sep 23, 2025).

What does Cognizant's new ES method mean for CTSH investors?

The announcement highlights potential efficiency gains in LLM training and strengthened IP portfolio, including two recent patents and a total of 61 U.S. patents.

Will Cognizant scale the ES fine-tuning to larger models and tasks?

Yes; the Lab stated its next step is to scale the ES method to fine-tune the largest available LLMs across complex tasks.

Who authored the ES fine-tuning research announced by Cognizant?

Authors include Xin Qiu, Yulu Gan, Conor F. Hayes, Qiyao Liang, Elliot Meyerson, Babak Hodjat and Risto Miikkulainen.