Cognizant’s AI Lab is shaking up how large language models (LLMs) are fine-tuned. The company announced a breakthrough technique that uses evolution strategies (ES)—a fundamentally different approach from traditional reinforcement learning (RL)—to train massive AI models more efficiently and at a fraction of the cost.
At the same time, Cognizant’s AI Lab added two new U.S. patents to its growing portfolio, underscoring its expanding influence in machine learning innovation.
Fine-Tuning Without the Reinforcement Learning Headache
The research, titled “Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning”, marks the first successful use of ES to fine-tune billion-parameter language models. Unlike reinforcement learning, which relies on massive amounts of labeled data and trial-and-error feedback loops, evolution strategies operate without gradients—directly exploring parameter space to find optimal model configurations.
The result? Fine-tuning that’s simpler, more stable, and up to 10 times faster, according to Cognizant’s internal benchmarks.
“Our approach not only uses less training data than reinforcement learning; it also makes the process more accurate, increasing the quality of work the AI can produce,” said Babak Hodjat, Chief AI Officer at Cognizant. “This has the potential to disrupt how the entire industry approaches fine-tuning.”
The Limits of Reinforcement Learning
Fine-tuning is what makes general-purpose models like GPT or Gemini specialized for real-world use—say, drafting contracts or writing code. But RL-based fine-tuning is resource-heavy, prone to overfitting, and can sometimes lead AIs to “game the system” rather than generate better answers.
Cognizant’s ES framework sidesteps these issues. It’s gradient-free, meaning it doesn’t depend on backpropagation, and it scales more gracefully as models grow larger. Early tests showed that the ES approach not only reduces computational overhead but also improves task alignment and consistency.
Since releasing its first ES fine-tuning code, Cognizant has achieved a 10x speed-up by integrating faster vLLM inference engines, setting the stage for fine-tuning the largest LLMs on the market.
AI Research With Real-World Impact
Cognizant’s research team—led by Xin Qiu, Yulu Gan, Conor Hayes, Qiyao Liang, Elliot Meyerson, Babak Hodjat, and Risto Miikkulainen—believes this work could redefine efficiency standards in model training.
As Miikkulainen, a UT Austin professor and Cognizant’s VP of Research, put it: “Obtaining millions of data points for model training is often unrealistic. These innovations enable effective model training with fewer examples, expanding the applicability of deep learning.”
Two New Patents Push the Frontier Further
Alongside its LLM breakthrough, Cognizant’s AI Lab secured two new U.S. patents—bringing its total to 61—that further advance AI’s practical capabilities:
- U.S. Patent No. 12,424,335 – AI-based Optimized Decision Making for Epidemiological Modeling: Uses neural networks to forecast epidemiological trends, combining multiple LSTM models for more accurate, real-world-constrained predictions.
- U.S. Patent No. 12,406,188 – Evolved Data Augmentation and Selection: Applies population-based search to automatically discover the best data augmentation methods, improving model performance on limited datasets.
Together, these patents highlight Cognizant’s focus on efficient AI for complex, data-scarce environments—from public health forecasting to commercial data optimization.
Why This Matters
The industry has been racing to make LLM training more affordable and sustainable. OpenAI, Anthropic, and Google all face spiraling costs as model sizes balloon and GPUs become scarce. Cognizant’s evolution strategies method could become a competitive alternative to reinforcement learning, particularly for enterprises fine-tuning domain-specific AI models without hyperscale resources.
As AI models grow more specialized, the ability to train smarter instead of bigger may define the next era of AI progress. And with Cognizant’s AI Lab now leading that conversation, the race to make large models more efficient—and more intelligent—just got a little tighter.
Power Tomorrow’s Intelligence — Build It with TechEdgeAI
