Call for Postdoctoral Researcher

Context: DEXTER4LLM is a collaborative R&D project under France-2023 aiming to build a sovereign-AI solution for structured extraction and automated analysis of complex industrial geographic documents (text, tables, figures, diagrams, plans, handwritten notes, etc.) [1], validated on high-impact industrial use cases (nuclear/mining & process engineering).

ECE leads the LLM Engineering (24 months) lot focusing on adaptation, optimization, and model merging to deliver frugal, high-performing geo-spatial (language / vision / multimodal) models for industrial workloads.

Mission: The post-doctoral will contribute to the scientific and engineering backbone, including:

• Document extraction and knowledge integration (knowledge graph / ontology).
• Domain adaptation / fine-tuning and model optimization.
• Model merging research & engineering: implement and evaluate merging recipes (e.g., DARE-TIES, SLERP-style interpolation, novel alternatives), with reproducible protocols and ablation studies.
• Benchmarking & metrics: accuracy/F1 for extraction, faithfulness (GraphRAG-style evaluation), and measurable frugality targets (compute reduction with environmental monitoring).
• Sustainability-aware evaluation tooling (carbon/compute tracking; reporting aligned with emerging standards).
• Research outputs: contribute to publications, open science where possible, and tech transfer with consortium partners.

The postdoctoral will work with the ECE team and in a consortium including major French enterprise.

References:

1. Ke, W., Zheng, Y., Li, Y., Xu, H., Nie, D., Wang, P., & He, Y. (2025). Large language models in document intelligence: A comprehensive survey, recent advances, challenges, and future trends. ACM Transactions on Information Systems, 44(1), 1-64.

https://www.ece.fr/

https://fr.wikipedia.org/wiki/%C3%89cole_centrale_d%27%C3%A9lectronique

PhD in Computer Science with ML/NLP/AI/GeoAI (or close field).

Strong Python + deep learning experience (PyTorch ecosystem).

Solid understanding of LLM training/fine-tuning (data pipelines, evaluation, failure modes).

Hands-on experience with at least one of:

• fine-tuning LLMs (LoRA/QLoRA),
• model compression/quantization,
• inference optimization / scalable serving,
• model merging / weight-space methods.

Ability to produce reproducible research (experiment tracking, clean repo practices) and produce high-impact research in Q1 journals / A level conference venues.