Stage 2027 AI/ML Engineer 100% R&D - Tabular foundation models

Latent Reasoning for Tabular Foundation Models: Performance Enhancement and Interpretability 

Ekimetrics is the European leader in data science, with over 500 data scientists and 1,000+ projects since 2006. With offices in Paris, London, New York, and Hong Kong, we conduct projects in over 50 countries across various sectors, including financial services, retail, telecom, healthcare, and more. 

Our mission is to help companies audit their data opportunities, enhance their analytical capital, and deploy actionable solutions that maximize their marketing and operational performance while reinvigorating business models. 

Our absolute focus is on delivering short-term gains while ensuring the long-term development of our clients' data capital. We are committed to offering the most advanced data science approaches and building ethical, sustainable AI practices. 

Key Figures 

• 20 years of experience in data science 

• 500+ data scientists, all consultants 

• 5 offices in Paris, London, New York, Hong Kong, and Shanghai 

• 350+ clients (CAC40, Fortune500) 

• $1M+ in profit generated for our clients since 2006 

• 1,000+ data science projects 

Within Ekimetrics, the Innovation Department works on AI research topics in collaboration with our industrial and academic partners. The department comprises several PhD experts in generative AI, deep learning, computer vision, time series, explainability, and causality. Each expert leads teams tasked with studying and building state-of-the-art methods and adapting them to specific business problems, developing new methodologies or algorithms to address identified challenges, and ensuring a seamless handover for integration within Ekimetrics' industrial ecosystem. 

A central focus of the Innovation Department is closing the gap between cutting-edge AI agents and the day-to-day reality of delivering data science at scale. The department is investing in agentic systems that accelerate the data science lifecycle while keeping data scientists firmly in control of quality, governance, and business relevance. This internship sits at the heart of that effort, combining applied AI research with production-grade engineering on Ekimetrics' modern data stack. 

Internship Objective: From Latent Reasoning to Better and More Understandable Tabular Models 

Tabular Foundation Models (TFMs) such as TabPFN, TabICL or TabDPT are transformers pre-trained on large collections of mostly synthetic datasets. They learn in context: the training data and the points to predict are given as input, and predictions come out of a single forward pass, with no dataset-specific training or tuning. They now rival gradient boosting on many benchmarks [3], which is changing how tabular modeling is practiced. 

In large language models, a growing line of research lets the model “reason” in its latent space rather than through generated tokens, using extra depth-wise computation, reused continuous hidden states or internal iterations. This latent reasoning can improve performance without longer text outputs [1, 2]. Our hypothesis is that a similar mechanism can improve TFMs: spending more latent compute at inference on hard or ambiguous cases could help them make better use of the in-context examples. Almost all existing work on latent reasoning targets natural language, and its transfer to tabular models, whose architecture, pre-training data and output differ profoundly, remains largely unexplored. 

Mechanistic interpretability is starting to open the black box, both for latent-reasoning language models and for TFMs themselves [4]. It lets us ask not only whether latent reasoning improves a TFM, but also what the model actually computes during these extra steps, and whether that computation is faithful and useful. The internship therefore has two goals: designing and evaluating latent reasoning mechanisms for TFMs, and using interpretability methods to analyze what they learn. 

The intern will work with senior researchers and data scientists. This internship offers an opportunity to contribute to an emerging research topic at the intersection of tabular modeling, latent reasoning and interpretability. 

Key Responsibilities 

The intern will: 

• Review the literature on latent reasoning (NLP) and tabular foundation models, building on recent surveys. 

• Identify and adapt latent reasoning mechanisms to TFM architectures (e.g. iterative depth-wise computation, hidden-state reuse). 

• Implement them on open-source TFMs and measure their effect on accuracy, calibration and robustness against strong baselines (gradient boosting, standard TFMs) at comparable compute. 

• Apply interpretability methods (representation analysis, probing, activation patching) to characterize the computation performed during latent reasoning. 

• Build an evaluation protocol on diverse tabular datasets, including realistic business cases. 

• Document the results and contribute to scientific publications. 

Profile 

• Currently in the final year of a Master’s or Engineering degree, with strong foundations in machine learning, statistics, and programming. 

• Good grasp of model validation: train/test splits, data leakage, baselines. 

• Strong analytical skills and the ability to work autonomously on an open-ended research problem. 

• Knowledge of transformers and language models; familiarity with latent reasoning or recurrent-depth models is a plus. 

• Strong Python and PyTorch skills, and a solid command of the standard data science toolkit (pandas/polars, scikit-learn, gradient-boosting libraries). 

• Fluency in English is a plus. 

• Nice to have: experience in interpretability, knowledge of TabPFN, TabICL or TabDPT, and solid software engineering practices (git, code review, testing). 

Benefits 

• Hands-on Research & Engineering Experience: Build a pioneering agentic system on real data science workflows, contributing to a tool with direct real-world impact. 

• Collaborative Learning: Engage with a team of data scientists, AI experts, and PhDs in a dynamic, research-driven environment within Ekimetrics' Innovation Department. 

• Skill Development: Gain expertise in LLM agents, the evaluation of AI systems, and production tooling on Databricks (Unity Catalog, MLflow). 

• Career Growth: Opportunity to transition into a full-time Data Scientist role in AI at Ekimetrics, or into a PhD program, upon successful completion of the internship. 

• Scientific Contribution: Opportunity to co-author and publish scientific papers in reputable AI and data science conferences and journals. 

• Networking: Collaborate with Ekimetrics' industrial and academic partners, enhancing your professional network in the AI research community. 

• Ethical and Sustainable AI: Learn to develop AI solutions that align with ethical, governed, and sustainable best practices, a core commitment of Ekimetrics. 

Duration: 6 months, Paris 

References 

[1] Zhu, R.-J. et al. (2025). A Survey on Latent Reasoning. arXiv:2507.06203. 

[2] Chen, X. et al. (2025). Reasoning Beyond Language: A Comprehensive Survey on Latent Chain-of-Thought Reasoning. arXiv:2505.16782. 

[3] Tabular Foundation Models: A Systematic Survey. [authors, year, link to complete] 

[4] Biloš, M. et al. (2026). A Mechanistic Study of Tabular Foundation Models. [link to complete]