Posted [contact details available in the full listing]
Job Information
- Organisation/Company: IMT Atlantique
- Research Field: Computer science » Other
- Researcher Profile: First Stage Researcher (R1)
- Positions: Postdoc Positions
- Application Deadline: 10 Oct 2026 - 00:00 (Europe/Paris)
- Country: France
- Type of Contract: Temporary
- Job Status: Full-time
- Hours Per Week: 38
- Offer Starting Date: 1 Jan 2027
- Is the job funded through the EU Research Framework Programme?: Not funded by a EU programme
Offer Description
Internationally recognized generalist engineering school of the IMT (Institut Mines-Télécom), leading French engineering school (Technological University), IMT Atlantique aims to support transitions, train responsible engineers, and use scientific excellence to serve teaching, research and innovation.
The position is based in Brest, in the BRAIN team (BRoader Artificial INtelligence) of Lab-STICC (UMRCNRS 6285). The team works on the foundations of machine learning (representations, learning from few examples, foundation models, signal processing) and on their applications, healthcare in particular. It recently developed REVE, a foundation model for electroencephalography pretrained on more than 25,000 subjects (NeurIPS 2025), and several of its PhD students work on discrete diffusion language models and on 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 on compute. In that regime, what a new example contributes no longer depends on its own quality alone, but on how it differs 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 the points and their similarity.
The project explores this question on two complementary levels: the diversity of the data, that is, which examples one annotates, keeps or presents to the model; and the diversity of the representations, that is, which features, contexts and models one exploits or combines. Results so far concern diversity-aware decoding for discrete diffusion language models, and the localisation of information within transformer representations.
This position addresses the second level, through the tools of mechanistic interpretability. Project page: https://bastienpasdeloup.github.io/endive/
MISSIONS
The main tasks of the position are:
- Define and evaluate diversity criteria in the representation space of foundation models, building on the tools of mechanistic interpretability.
- Study the relationship between the diversity of training data and the diversity of the internal mechanisms models acquire.
- Contribute to the scientific output and dissemination of the project, and to the supervision of its interns.
ACTIVITIES:
- Diversity criteria in representations:
- Train and analyze sparse autoencoders on the activations of foundation models, in order to extract interpretable features.
- Build similarity kernels between those features and evaluate, by means of determinantal point processes, whether a diverse subset covers them more compactly than individual importance alone.
- Measure the effect of these criteria on downstream tasks in the low-annotation regime, and compare the network depths at which representations are read.
- Data diversity and mechanism diversity:
- Compare the features recovered by sparse autoencoders trained on models fed different data diets, and quantify their overlap.
- Assess the stability of those features across training runs, so as to separate reproducible mechanisms from optimisation artefacts.
- Relate these measurements to the curation procedures studied in the project, in particular for discrete diffusion language models.
- Scientific output and project life:
- Writes up and submits the results obtained to international conferences and journals.
- Releases the code and experimental protocols required to reproduce the results.
Where to apply
Requirements
- Research Field: Computer science » Other
- Education Level: PhD or equivalent
Skills/Qualifications
Minimum education and/or experience required:
🎓 PhD obtained less than 3 years before date of hire in machine learning, computer science, signal processing or applied mathematics.
Essential skills, knowledge and experience:
✔️ Solid command of deep learning and of its mathematical foundations.
✔️ Fluent practice of Python and of a deep learning framework, preferably PyTorch.
✔️ Experience with transformer architectures, and the ability to instrument their internal representations.
✔️ Autonomy in running GPU experiments, including on shared computing infrastructure.
🇬🇧 Publications in international conferences or journals of the field, and a very good level of scientific English, written and spoken.
✔️ Familiarity with mechanistic interpretability (sparse autoencoders, linear probes, circuit analysis), with determinantal point processes or sampling methods more generally, with discrete diffusion language models, or with multi-GPU computing on a SLURM-like scheduler: appreciated, but none of these is required.
Abilities and skills:
✔️ Scientific autonomy and an appetite for open questions, the position having a deliberately exploratory part.
✔️ Experimental rigor and attention to reproducibility.
✔️ A taste for collective work: the position interacts directly with several PhD students of the project.
✔️ Good writing skills, and the ability to present results to a non-specialist audience.
OTHER INFORMATION
• Command of French is not required: English is the working language of the team as soon as one of its members is not a French speaker.
• Publications resulting from the project are deposited in open access, in line with the commitments made to the ANR.
Additional Information
Benefits
WHY JOIN US :
👉 Friendly working environment in a small, dynamic team
👉 Stimulating innovation ecosystem (startups, students, research, companies)
👉 Collaboration with renowned research organizations
👉 Collaboration with industry
THE PLUS
🖥️Partial working from home possible
🥗 Collective catering on site
🚌 Public transport paid for
🚴♂️Sustainable mobility package (for carpooling or cycling)
👨👩👧👦 Family supplement
💶 Wide range of social benefits
🌴 Numerous vacations
Selection process
Deadline for application : 10/10/2026
Start of the contract : 01/01/2027
Interviews : real-time
Website for additional job details: https://institutminestelecom.recruitee.com/l/en/o/post-doctorat-en-interpretabilite-mecaniste-et-diversite-des-representations
Work Location(s)
- Number of offers available: 1
- Company/Institute: IMT Atlantique
- Country: France
- City: Brest
- Postal Code: 29200
- Street: 655 avenue du technopôle 29280 Plouzané
Contact
- City: Brest
- Website: https://www.imt.fr/
- Street: 655 avenue du technopôle 29280 Plouzané
- Postal Code: 29200
- E-Mail: bastien.pasdeloup@imt-atlantique.fr
This is a Preview Listing…
You must sign in to see the full job description, and to apply.
Manage / Upgrade this job to a Full Job Listing.
Find Your Best Opportunity
Tell them AcademicJobs.com sent you!