18 MONTH POST-DOC ON RAG, GAR AND GENERATIVE INFORMATION RETRIEVAL Post-doc opening in the area of Generative Information Retrieval at the AI Institute MIAI (https://miai-cluster.univ-grenoble-alpes.fr/english/) and Laboratoire d'Informatique de Grenoble (https://www.liglab.fr/en) France. OVERALL PROJECT SUMMARY The goal of the research project is to combine Retrieval-Augmented Generation (RAG), Generation-Augmented Retrieval (GAR) and Generative Information Retrieval (GenIR) to develop a state-of-the-art information retrieval system. RAG emerged in 2020 as a way to increase the quality of LLM outputs by reducing the hallucinations inherent to these models ([1]). In RAG, instead of directly asking an LLM to answer a user question, one first passes the question to an IR system which retrieves passages relevant to it (retrieval part); the LLM then produces the final answer, using both the question and the retrieved passages as input (generation part). The retrieved passages thus help the LLM to find the correct answer and avoid hallucinating. RAG also has the advantage of providing links to some of the sources of the information used to provide the answer ([2]). Most recent LLMs integrate (variants of) RAG, which have become the state of the art in terms of generation. GAR can be seen as the counterpart of RAG inasmuch as one first augments a given query with answers provided by an LLM (generation part) prior to retrieve passages using both the original query and the LLM-generated answers (retrieval part). Using the additional information provided by the LLM for the retrieval step increases the quality of the retrieved passages in a GAR architecture, especially when a RAG system is used to generate the first answers from the query [3]. Both RAG and GAR however suffer from two main problems: (a) the potential mismatch between the generation and the retrieval parts may prevent developing a fully integrated model, as gradient updates may not propagate from the end task to all components, and (b) the quality of retrieval, respectively generation, may impact the generation, respectively retrieval, outcome. Recently, Tang et al. ([4]) proposed to use the same generative decoder-only model for the retrieval and generation steps of RAG and GAR, combining the retrieval and generation losses intoa single loss to obtain fully end-to-end models, thus addressing problem (a). They furthermore experimentally showed that using a GAR model for the retrieval step of a RAG system increases the quality of the generated answers and leads to state-of-the-art generative models, thus addressing problem (b) for RAG; they however left problem (b) unanswered for GAR. Our main objective in this project is to design a theoretical and experimental framework to train and evaluate state-of-the-art GAR-based IR systems. For that, we will rely on a RAG architecture for the generation part of GAR, and GenIR models for all the retrieval parts, with the same underlying decoder-only LLM. Proposals will be evaluated on state-of-the art IR benchmarks. References [1] P. Lewis, E. Perez, A. Piktus, F. Petroni, V. Karpukhin, N. Goyal, H. Küttler, M. Lewis, W. -T. Yih, T. Rocktäschel, S. Riedel, D. Kiela. Retrieval-augmented generation for knowledge-intensive NLP tasks, NeurIPS 2020. [2] I. Nematov, T. Kalai, E. Kuzmenko, G. Fugagnoli, D. Sacharidis, K. Hose, T. Sagi. Source Attribution in Retrieval-Augmented Generation, CoRR abs/2507.04480 (2025). [3] D. Arora, A. Kini, S. R. Chowdhury, N. Natarajan, G. Sinha, A. Sharma. GAR-meets-RAG Paradigm for Zero-Shot Information Retrieval, CoRR abs/2310.20158 (2023). [4] Y. Tang, R. Zhang, J. Guo, M. de Rijke, Y. Fan, X. Cheng. Boosting etrieval-Augmented Generation with Generation-Augmented Retrieval: A Co-Training Approach, SIGIR 2025. PROJECT LEADERS Eric Gaussier https://ama.liglab.fr/~gaussier/ Philippe Mulhem https://lig-membres.imag.fr/mulhem/ APPLICATION REQUIREMENTS: Completed doctoral degree in the area of computational linguistics, natural language processing, or information retrieval with a strong machine learning experience. APPLICANT QUALIFICATION PROFILE: - Solid mathematical and programming skills - Interest in inter-disciplinary research, in particular between multiple areas of computer science - Ability to work independently as well as in teams - High level of English - Relevant Publications HOSTING INSTITUTION One of the major research-intensive French universities, Univ. Grenoble Alpes enjoys an international reputation in many scientific fields, as confirmed by international rankings. The dynamic ecosystem, grounded on a close interaction between research, education and companies, has earned Grenoble to be ranked as the 5th most innovative city in the world. Surrounded by mountains, the campus benefits from a natural environment and a high quality of life and work environment. With 7000 foreign students and the annual visit of more than 8000 researchers from all over the world, Univ. Grenoble Alpes is an internationally engaged university. A personalized Welcome Center for international students, PhDs and researchers facilitates your arrival and installation. The MIAI Cluster (Multidisciplinary Institute in Artificial Intelligence) is one of the nine French AI Clusters dedicated to artificial intelligence research, innovation, and education. Coordinated by Université Grenoble Alpes and developed in collaboration with Université Clermont Auvergne and Université Savoie Mont Blanc, the institute brings together researchers, companies, public institutions, startups, and students to address the major scientific, technological, and societal challenges of AI. Lastly, Grenoble Informatics Laboratory (LIG) is one of the largest laboratories in Computer Science in France. It is structured as a Joint Research Center (French Unité Mixte de Recherche - UMR) founded by the following institutions: CNRS, Grenoble Institute of Technology (Grenoble INP), Inria Grenoble Rhône-Alpes, Grenoble Alpes University. The mission of LIG is to contribute to the development of fundamental aspects of Computer Science (models, languages, methodologies, algorithms) and address conceptual, technological, and societal challenges. WHAT WE OFFER: - Intellectually-motivating environment with opportunities for inter-disciplinary work. - Accesses to GPU infrastructures (LIG, Grid 5000, Jean Zay, Adastra), to run experiments. - Fixed-term employment contract (18 months). - Funding: The PostDoc is funded by the MIAI Cluster (Multidisciplinary Institute in Artificial intelligence of the Grenoble Alpes University). - Gross salary: from 2300¤/month to 2600¤/month depending on the candidate's experience. HOW TO APPLY: Applicants should send a complete CV, including a list of publications, to Philippe Mulhem (Philippe.Mulhem@imag.fr) and Éric Gaussier (eric.gaussier@univ-grenoble-alpes.fr) before 15 July 2026. The successful candidate is expected to start by January 1, 2027, at the latest.