ANITI Toulouse ’s Post

🚀 Postdoc Opportunity : Fairness metrics for open weight generative models 🌟 The CALM and TRIAL research chairs of ANITI Toulouse are offering a postdoctoral position to tackle one of AI’s most pressing challenges: developing innovative fairness metrics and correction algorithms for generative models. This role provides a unique opportunity to engage in cutting-edge research, analyzing bias in models like LLaMA and Mistral, and contributing to methods that enhance fairness in AI systems. Contact : Jean Michel Loubes Subject : The primary aim of the internship is to gain a deep understanding of what constitutes bias in the context of generative AI. Initially, the intern will conduct a comprehensive literature review to assess the current state of knowledge on bias in generative models. Based on this review, the intern will then work to develop new bias metrics. Classical approaches to fairness and privacy generally assume access to the training data. However, with open-source large language models and diffusion models, only the model weights are available (and for closed models, only the inference results). Additionally, these models are general-purpose, meaning fairness concerns are relevant only for specific applications. The challenge, then, is to consider fairness metrics for large generative models when only the weights are accessible. One approach is to examine the distributions and attention masks in the latent space in relation to a target sensitive variable. Similar approaches can also apply to differential privacy, particularly in detecting whether protected content was part of the training set. The intern will compare the proposed metrics across various open-source LLM models, such as LLaMA and Mistral, and contribute to developing bias-correction methods, offering practical solutions to improve fairness in generative models. Candidate Profile: We are looking for a candidate with a strong background in machine learning and computer science. Experience with generative AI algorithms—such as transformers or diffusion models— is also important. About the Research Group: You will be working within the chairs CALM  and TRIAL (Trustworthy AI in Machine Learning) research chair, a multidisciplinary team that brings together experts in machine learning, optimization, and computer science to ensure AI systems are reliable, fair, and transparent. Internship Duration: The Post  will last 1 year with a possible 1 year extension, providing an excellent opportunity to engage in cutting-edge research while working alongside leading experts in from ANITI and the ANR Regulia. Location : Toulouse (ANITI) This internship offers the chance to contribute to crucial advancements in ensuring robustness and fairness in AI, particularly in the growing field of generative models. If you are passionate about the intersection of AI, ethics, and innovation, this is a fantastic opportunity to expand your expertise and make a meaningful impact.