PhD "Detection and mitigation of LLM summary bias for Enterprise services" M/F

about the role

Your role is to conduct a PhD thesis on the detection and mitigation of biases in summaries generated by LLMs in Enterprise services, for Artificial Intelligence (AI) applications such as meeting summaries or verbatim reports within an organization, client interaction syntheses, and summaries of unstructured interpersonal exchanges via communication and collaboration services. This is not merely about summarizing exchanges but producing a synthesis that is useful for its recipient, whose needs may vary depending on their profession or objectives.

However, the production of this synthesis can be tainted by biases that may lead to inequalities, inaccuracies, or harm, whether intentional or not. These biases arise from various factors, such as training data, design choices, or interactions with users. With the emergence of self-supervised pretrained large language models, it is no longer possible to measure correlations between input signals, algorithmic choices, and observed biases. Regarding the detection of these biases, only an observation-based approach is feasible, relying on the construction of metrics with evaluations conducted by humans and/or language models (LLMs as Judges).

The main steps of the PhD thesis will therefore be:

1. To study and develop a suitable methodology and evaluation protocols, with a focus on modularity and generalizability (multi-LLM, multi-metric, and multi-use cases), and to implement a benchmark that incorporates them.

2. To conduct tests based on this benchmark by defining dedicated metrics.

3. To identify bias mitigation strategies by comparing approaches with or without additional model training.

One of Orange's strengths in this work is the availability of private datasets that have not contaminated the models during either the pre-training or alignment phases. Public datasets will also be used. The metrics will be developed through progressive iteration, as biases manifest statistically. It will be necessary to go beyond testing and extend to the search for heuristic rules regarding the correlation between the various stages of construction and use of LLMs with the presence of biases. Finally, a suitable method will be constructed to correct these biases by identifying the most appropriate alignment strategies: using supervised learning to fine-tune a pretrained model, or through reinforcement learning. The proposed solution will be compared to bias mitigation methods performed without additional training (auto-diagnosis, prompt engineering, and roleplay).

Context: https://hellofuture.orange.com/fr/ne-pas-reproduire-prejuges-et-erreurs-humaines-dans-les-llms-comment-faire/