“Artificial intelligence for the Sciences” (AI4theSciences) is an innovative, interdisciplinary and intersectoral PhD programme, led by Université Paris Sciences et Lettres (PSL) and co-funded by the European Commission. Supported by the European innovation and research programme Horizon 2020-Marie Sklodowska-Curie Actions, AI4theSciences is uniquely shaped to train a new generation of researchers at the highest academic level in their main discipline (Physics, Engineering, Biology, Human and Social Sciences) and master the latest technologies in Artificial Intelligence and Machine Learning which apply in their own field. 26 doctoral students will join the PSL university's doctoral schools in 3 academic cohorts to carry out work on subjects suggested and defined by PSL's scientific community. The 2023 call will offer up to 4 PhD positions on 10 PhD research projects. The candidates will be recruited through HR processes of high standard, based on transparency, equal opportunities and excellence.

1. Context/Motivation for the project

Large Language Models (LLMs) such as BigScience’s BLOOM and OpenAI’s ChatGPT are groundbreaking AI systems and are among the fastest new technologies to be so widely adopted in terms of the number of startups they have fostered and their rapid integration into daily use technologies such as chatbots, web search and writing assistance softwares etc. LLMs are autoregressive language models (Schuster & Hegelich, 2022) based on the transformer architecture that use deep learning (Hegelich, 2017) to produce human-like text. They have been called Foundation Models in reference to their zero/one/few-shot generalization capabilities (Model doesn’t need training on specific tasks but can perform a wide range of tasks based on pretrained associations). They pick up on elaborate languagebased behaviors in response to given prompts and approximate the complexity of human natural language to a surprising degree. A capacity for natural human language also makes them uniquely positioned for human-centered interactive AI applications.  

However, as a number of recent studies, focussed on bias and fairness, have revealed (Caliskan et al., 2017; Papakyriakopoulos, Hegelich et al., 2020; Bender et al., 2021; Weidinger et al., 2021), such models, once developed and deployed in the real world, show emergent properties that often deviate from the intent and values of their developers. Since the vectorized representations in language models are derived from billions of examples of real-world language use, which in turn are derived from meaning ascribed to them by humans, it is not surprising that LLMs can absorb, abstract and amplify human characteristics (Caliskan et al., 2017; Bender et al., 2021). However, as these characteristics are emergent and were not explicitly introduced by the developers, they are unknownunknowns and can only be uncovered by a broad-based methodical examination of the output.

2a. Scientific Objectives

This thesis would be aimed at producing a wider characterization of what aspects, among the very wide range of psycho-social-cultural characteristics that underlie human behavior, are captured and reflected in the behavior exhibited by LLMs. To this end, we will probe textual content generated by base versions of LLMs in response to experimenter’s prompts, with a range of psychological, social and cultural constructs and establish which effects are reliable and internally consistent.  

We will also use this lens to study how these specific features are used in context (not just in the initial vectors but in the actual representations calculated during processing). This is a key objective to understand how encoding is actually used and how biases could be amplified by the algorithms used. This will be done through the implementations of dynamic methods, esp. ablation techniques that are now widely used in the domain (see next sections).  

Although the thesis would mainly involve empirical studies to uncover what (and how) social, cultural and psychological dimensions are encoded in LLMs, a reflection on the consequences of this encoding will also be considered, opening perspectives for subsequent research. In the last part, we will define an ethical framework under which elements in an AI model that mirror human-like behaviors (personality traits, cultural preferences etc.) could be situated among broader concerns of AI ethics. These questions hold direct relevance to the problem of AI value alignment. What does it mean for an

AI model to have an extraverted personality or to reflect specific moral and cultural values? Would such emergent preferences in base models change an AI system’s understanding of human instructions?

2b. Methodology

For text generation, we will primarily be working with open-source Large Language models such as BLOOM (BigScience Workshop et al., 2022) as they allow for unfiltered examination of the base model.  

On the behavioral side, we would utilize a range of established social scientific measures such as empirical measures of Personality and Ethical Values. We will use advanced NLP (Natural Language Processing) techniques for automatic psychographic assessment from text data (Tausczik & Pennebaker, 2010) as well as assessments rated by human users after guided interaction. Prof. Hegelich’s team at TUM will provide a model (based on Google’s BERT) that can estimate personality traits from text samples of human subjects. Other required NLP measures will be collaboratively developed in the course of the thesis as needed. In addition, various customizable LLM variables (temperature, model size, multilinguality) will be tested for interaction with emergent behaviors.  

A central quest of probing is to uncover how pre-trained models encode linguistic properties within their representations. Following previous experiments (see Lasri et al., 2022), in the course of the thesis, we will try to tackle how encoding is used by the model, introducing a usage-based probing setup. Thanks to recent techniques like causal probing, it is now possible to remove a given property by intervening on the model’s representations. We contend that, if an encoding is used by the model, its removal should harm the performance on the chosen behavioral task. Past experiments were mainly about linguistic properties (like subject-verb agreement). We will extend these experiments to get a better understanding of how information (especially along social, cultural and psychological dimensions) is encoded in the network.  

2c. Expected Results

The thesis will lead to both scientific and practical results:  

• On the scientific side, the thesis will lead to a better characterization of social, cultural and psychological dimensions in LLMs. This is important especially as LLMs are soon expected to power wide-ranging daily use technologies. Some of these dimensions may need to be preserved and activated depending on the profile of the LLM’s user (e.g. cultural dimensions) but others may require mitigation efforts (e.g. bias, non-preferred behaviors).  
• On the practical side, the thesis will lead to proposals about how to practically take these different dimensions into account. The thesis would depart from the current approach aimed at just removing biases from language models, by providing a more complex image of different traits that could be adjusted depending on the profile of the user. A full characterization of this kind would also in turn be useful to improve control over LLM output.

3a. 3i Characteristics of the Thesis

The project is interdisciplinary by default in both its subject and methodologies. The thesis will examine generative AI models in the domain of Language, namely, Large Language Models (also called Foundation models) both in terms of their structure and output.  

This examination will be done in light of knowledge, perspectives and methodologies from a range of social science disciplines – psychological, sociological and cultural studies. In addition to such behavioral aspects, we will also examine LLMs by going ‘inside’ the network, using techniques such as ablation. We will also examine effects of multilinguality (e.g. a French LLM vs. German LLM) and model size (in terms of parameters) on the observed behaviors. International

The thesis involves a close active collaboration between researchers at Prof. Poibeau’s group at the Ecole normale supérieure-PSL and Prof. Hegelich’s group at the Technical University of Munich, Germany.

This research aims at exploring how AI through LLMs impact society. This is a growing topic of interest within PSL, especially with respect to the chair Abeona-ENS-OBVIA on AI and social justice (Scientific advisor: T. Poibeau). The PhD will contribute to fostering the reflections in these domains (bias, social justice, etc.) within PSL and beyond.  

3b. Expected Impact

LLMs are expected to influence every aspect of our society that involves the written word, ranging from education to journalism and politics. A broad understanding of LLM behavior is thus both urgent and critical.  

The behavioral aspects uncovered in this project will, first, be influential in developing the norms underlying the deployment of LLMs, which will in turn be relevant to the ongoing legislative and regulatory discussions on AI governance.

Second, LLMs are one of the most active areas in AI research. Characterization of novel behaviors reflected in LLM output can form an important reference for developers and researchers of such systems.

3c. Feasibility of the Thesis in 3 years

We are confident that this PhD is feasible in a strict 36-month time period. Lattice has a strong expertise in NLP and especially in the exploration of the content of  LLMs. TUM, on the other hand, has a strong experience and other ongoing projects in the exploration of social and cultural dimensions within these models. The successful candidate will benefit from the experience of both teams and will progress according to a clear and reasonable schedule defined at the beginning of the thesis and regularly adjusted if necessary.  

References

Bender, E. M., Gebru, T., McMillan-Major, A., & Shmitchell, S. (2021). On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, 610–623.

BigScience Workshop, Scao, T. L., Fan, A., Akiki, C., Pavlick, E., Ilić, S., Hesslow, D., Castagné, R., Luccioni, A. S., Yvon, F., Gallé, M., Tow, J., Rush, A. M., Biderman, S., Webson, A., Ammanamanchi, P. S., Wang, T., Sagot, B., Muennighoff, N., … Wolf, T. (2022). BLOOM: A 176B-Parameter Open-Access Multilingual  Language           Model(arXiv:2211.05100). arXiv. https://doi.org/10.48550/arXiv.2211.05100

Caliskan, A., Bryson, J. J., & Narayanan, A. (2017). Semantics derived automatically from language corpora          contain       human-like         biases. Science,               356(6334),          183–186. https://doi.org/10.1126/science.aal4230

Hegelich, S. (2017). Deep learning and punctuated equilibrium theory. Cognitive Systems Research, 45, 59–69. https://doi.org/10.1016/j.cogsys.2017.02.006

Karim Lasri, Tiago Pimentel, Alessandro Lenci, Thierry Poibeau, Ryan Cotterell: Probing for the Usage of Grammatical Number. ACL (1) 2022: 8818-8831

Papakyriakopoulos, O., Hegelich, S., Serrano, J. C. M., & Marco, F. (2020). Bias in word embeddings. Proceedings of the 2020 Conference on Fairness, Accountability, and Transparency, 446–457.

Schuster, C. M., & Hegelich, S. (2022). From BERT’s Point of View: Revealing the Prevailing Contextual Differences. Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics.

Tausczik, Y. R., & Pennebaker, J. W. (2010). The Psychological Meaning of Words: LIWC and Computerized Text Analysis Methods. Journal of Language and Social Psychology, 29(1), 24–54. https://doi.org/10.1177/0261927X09351676

Weidinger, L., Mellor, J., Rauh, M., Griffin, C., Uesato, J., Huang, P.-S., Cheng, M., Glaese, M., Balle, B., Kasirzadeh, A., Kenton, Z., Brown, S., Hawkins, W., Stepleton, T., Biles, C., Birhane, A., Haas, J.,Rimell, L., Hendricks, L. A., … Gabriel, I. (2021). Ethical and social risks of harm from Language Models. ArXiv:2112.04359 [Cs]. http://arxiv.org/abs/2112.04359   

Thesis supervisors

Main PhD Supervisor: Thierry Poibeau (Lattice, ENS-PSL, Prairie Fellow)

Second PhD Supervisor: Prof. Simon Hegelich (Political Data Science, Technical University Munich)

PSL University PSL

Created in 2012, Université PSL is aiming at developing interdisciplinary training programmes and science projects of excellence within its members. Its 140 laboratories and 2,900 researchers carry out highlevel disciplinary research, both fundamental and applied, fostering a strong interdisciplinary approach. The scope of Université PSL covers all areas of knowledge and creation (Sciences, Humanities and Social Science, Engineering, the Arts). Its eleven component schools gather 17,000 students and have won more than 200 ERC. PSL has been ranked 36 in the 2020 Shanghai ranking (ARWU).