Colombes, France – October 19th, 2026 – ALE International, operating under the Alcatel-Lucent Enterprise brand name, welcomes the Best Paper Award received by Estelle Zheng at TALN 2026, the leading French-language conference on Natural Language Processing (NLP). Doctoral candidate at LORIA, (Laboratoire Lorrain de Recherche en Informatique et ses Applications), Estelle Zheng is conducting a CIFRE doctoral thesis (Industrial Agreement for Training through Research) in partnership with Alcatel-Lucent Enterprise, bringing together academic research and industrial challenges. In the award-winning paper, she demonstrates that a large language model can be specialized while reducing peak memory requirements by approximately half, with no significant loss in performance. This approach opens up new opportunities for organizations seeking to adapt AI models while maintaining control over their data and infrastructure, even with limited computing resources.
Estelle Zheng’s journey: from data science to artificial intelligence research
After earning a Bachelor's degree in Electronic Engineering from Sorbonne Université, Estelle Zheng completed a Master's degree in Applied Mathematics and Data Science at the Institut Polytechnique de Paris. In 2023, as generative AI was becoming widely accessible to the public, she decided to pursue doctoral studies to better understand and master this technology.
PhD opportunities on large language models were still quite rare in France in 2023 when I started looking, just before the generative AI boom. Christophe Cerisara, a CNRS researcher specializing in LLMs and Natural Language Processing and a lead researcher at LORIA, contacted me to propose this topic. Pursuing this PhD is a tremendous opportunity. I am fortunate to have supervisors at both LORIA and Alcatel-Lucent Enterprise who give me a great deal of freedom and time to explore, investigate, test, fail, understand why, and ultimately build knowledge.
Award-winning research addressing a key industrial challenge
The award-winning paper « Yes to Scale, No to Memory: Lightweight Fine-Tuning of LLMs with Side Networks » (« Oui à l'Échelle, Non à la Mémoire : Affinage Léger des LLMs par Réseaux Latéraux ») is part of Estelle Zheng’s doctoral work entitled “Fine-Tuning Large Language Models for Planning and Action via APIs”, conducted at LORIA under the supervision of Christophe Cerisara and in partnership with Alcatel-Lucent Enterprise.
The paper was presented at TALN 2026, held in Nantes this summer. It addresses a key issue: for enterprise deployment, a general-purpose large language model (LLM) typically needs to be fine-tuned to a specific domain or use case. However, this process requires substantial computing resources, including high-end graphics processing units (GPUs), which are costly and not always readily available.
In her research, Estelle Zheng explores a still relatively under-investigated architecture called “Ladder”, which consists of adding a small side network to the main model. Her results show that this approach achieves performance improvements comparable to QLoRA, one of the most widely used fine-tuning methods, while reducing peak memory consumption by approximately 50%. She also introduces a variant called, « xLadder », which further develops the side-network concept.
How Alcatel-Lucent Enterprise uses research to stay at the forefront of innovation
In 2024, Alcatel-Lucent Enterprise launched an internal project aimed at simplifying the configuration of its unified communications systems through intelligent models. At a time when generative AI was still emerging, the company chose to draw on academic research to explore these technologies and address this challenge. Discussions with Christophe Cerisara’s team at LORIA led to the creation of Estelle Zheng’s CIFRE doctoral program.
Since then, the project has evolved in step with advances in artificial intelligence. While the original applications have changed, the underlying approach remains the same: starting from concrete industrial needs to advance knowledge and identify new technological opportunities. This collaboration enables Alcatel-Lucent Enterprise to benefit from research expertise closely aligned with state-of-the-art developments in a rapidly evolving field.
This initiative continues ALE’s long-standing collaborations with the academic community. The company previously supported a doctoral thesis in 2014 and currently participates in several European research and innovation programs, including Cortex2, focused on AI-powered telepresence, and ShareSpace, which examines the societal impact of extended reality. Both projects are coordinated by the German research center DFKI under the Horizon Europe Framework.
In a technological environment that evolves at an incredible pace, relying on research helps us stay ahead. When we designed this project two years ago, we could not have imagined how much, or how quickly, it would evolve. Working with someone like Estelle Zheng, who continuously monitors the latest advances, allows us to remain at the state of the art. Her research and close collaboration with our teams help disseminate knowledge internally. We strongly believe that maintaining close ties with academic research provides valuable perspective and enriches our projects.
Estelle Zheng will also present a new paper based on her doctoral research at EMNLP 2026, the leading international conference on Natural Language Processing, which will take place in Budapest at the end of October 2026. Her paper was accepted among 17,669 submissions from around the world.