Post-doctoral position in Mechanistic Interpretability and Representation Diversity

A major generalist engineering school of IMT-Institut Mines-Télécom, the leading group of engineering schools in France, IMT Atlantique aims to support transitions, train responsible engineers, and put scientific and technical excellence at the service of teaching, research, and innovation.

The position is based in Brest, within the BRAIN (BRoader Artificial INtelligence) team of Lab-STICC (UMR CNRS 6285). The team works on the foundations of machine learning (representations, frugal learning, foundation models, signal processing) and their applications, particularly in health. It has recently developed REVE, a foundation model for electroencephalography pre-trained on over 25,000 subjects (NeurIPS 2025), and is interested in discrete diffusion language models and mechanistic interpretability.

The ENDIVE project studies what diversity can bring to machine learning when the budget is on the number of annotated examples rather than computational power. In this regime, what a new example brings no longer depends solely on its own quality, but on what distinguishes it from those already seen. The technical entry point of the project is sampling with diversity guarantees, and in particular Determinantal Point Processes (DPPs), whose kernel matrix encodes both the relevance of points and their similarity.

The project explores this question on two complementary levels: data diversity, i.e., the choice of examples that are annotated, kept, or presented to the model; and representation diversity, i.e., the choice of descriptors, contexts, and models that are exploited or

combined. The first results concern the diversified decoding of discrete diffusion language models and the localization of information in transformer representations.

The proposed position focuses on the second level, approached through the tools of mechanistic interpretability. Project page: https://bastienpasdeloup.github.io/endive/

MISSIONS

The main missions of the position are as follows:

1. Define and evaluate diversity criteria in the representation space of foundation models,

leveraging the tools of mechanistic interpretability.

2. Study the link between the diversity of training data and the diversity of internal mechanisms

that models acquire.

3. Contribute to the scientific production and dissemination of the project.

ACTIVITIES:

1. Diversity criteria in representations:

• Train and analyze sparse autoencoders on the activations of foundation

models to extract interpretable descriptors.

• Construct similarity kernels between these descriptors and evaluate, using determinantal

point processes, whether a diverse subset provides a more compact coverage than individual importance alone.

• Measure the effect of these criteria on downstream tasks in low-annotation regimes, and compare the

network depths at which representations are read.

2. Data diversity and mechanism diversity:

• Compare the descriptors found by sparse autoencoders trained on

models fed with different data regimes, and quantify their overlap.

• Evaluate the stability of these descriptors from one training session to another, in order to distinguish reproducible

mechanisms from optimization artifacts.

• Relate these measurements to the curation procedures studied in the project, particularly for discrete diffusion language models.

3. Scientific production and project life:

• Write and submit the obtained results to international conferences and journals.

• Publish the code and experimental protocols necessary for the reproducibility of the results.

Minimum educational level and/or experience required:

🎓 PhD obtained less than 3 years before the start date, in machine learning, computer science, signal processing, or applied mathematics.

Essential skills, knowledge, and experience:

✔️ Solid mastery of deep learning and its mathematical foundations.

✔️ Fluent in Python and a deep learning framework, preferably PyTorch.

✔️ Experience working with transformer-type architectures, and ability to instrument their internal representations.

✔️ Autonomy in conducting experiments on GPUs, including on shared computing infrastructure.

🇬🇧 Publications in international conferences or journals in the field, and a very good level of scientific English, both written and spoken.

Desirable skills, knowledge, and experience:

✨ Knowledge of mechanistic interpretability: sparse autoencoders, linear probes, circuit analysis.

✨ Familiarity with determinantal point processes, or more generally with sampling methods and random linear algebra.

✨ Interest in discrete diffusion language models.

✨ Experience with multi-GPU distributed computing and a SLURM-type scheduler.

Abilities and aptitudes:

✔️ Scientific autonomy and interest in open questions, with an assumed exploratory component to the position.

✔️ Experimental rigor and attention to reproducibility.

✔️ Enjoyment of teamwork: the position involves direct interaction with several PhD students on the project.

✔️ Writing skills, and ability to present results to a non-specialist audience.

ADDITIONAL INFORMATION

• Proficiency in French is not required: English is the team's working language as soon as one of its members is not a French speaker.

• Publications resulting from the project are deposited

WHY JOIN US:

👉 Exceptional work environment

👉 Stimulating innovation ecosystem (startups, students, research, companies)

👉 Collaborations with renowned research organizations

👉 Collaboration with the industrial sector

BONUSES:

🥗 On-site collective catering

🏃‍♀️ Leisure/sports offerings

🚌 Public transport reimbursement

🚴‍♂️ Sustainable mobility package (for carpooling or cycling commutes)

👨‍👩‍👧‍👦 Family allowance

💶 Wide range of social benefits

🖥️ Partial remote work possible

🌴 Numerous holidays

 REMUNERATION GUIDELINES:

💰 Indicative remuneration: €35,800 gross per year, depending on profile and experience (reduced employee contributions in the public sector).

 For any information*:

• On the content of the position: Bastien PASDELOUP – Associate Professor: On administrative/HR aspects: Mélissandre Morvan – Recruitment Assistant:

  Application deadline: 10/10/2026

  Contract start date: 01/01/2027

  Interviews: on a rolling basis

   

   

  Legal Mentions [1]:

  • Nature and duration of the contract: 18-month fixed-term contract under public law or detachment on contract for civil servants, renewable.

  • Job location: IMT Atlantique Brest campus - 655 Av. du Technopôle, 29280 Plouzané

  • For internal use:

    • Category – Job type: II P

  • Positions offered for recruitment are open to all, with, upon request, accommodations for candidates with disabilities.

  • Job open to public service officials and/or contract employees.

  • All applications may be subject to an administrative investigation.

  Find Jobs in France on Arbeitnow