Lucas Biéchy

LucasBiéchy✺

Lucas' memoji waving hello

I explore what language models can do and how they go wrong, from evals to theory.

PhD researcher at Inria in the PETSCRAFT team, fellow at MARS V and BlueDot Impact alumnus.

Read my research Email meView my CV

The short version

Selected paper

Empirical · Under review

PrivacySkills

30%

of runs where LLM agents dig into confidential sources even though a harmless option exists. Privacy labels on the skills plus a system instruction roughly halve it.

Selected paper

Theory · Preprint

Testing a Learner Changes It

7×

fewer rounds than random probes to catch a mildly biased self-grading agent (57 vs 424), with a lower bound showing no adaptive audit can do better.

MARS V · summer 2026

7 weeks

From the MARS V kick-off in mid-July to a paper accepted at the BiAlign Workshop @ NeurIPS: From Observation to Intervention, on identifying latent user bias in human-AI interaction.

Upcoming project

Theory → practice

Testing a Learner Changes It, applied to Hermes

Hermes learns from its own judgments of success, so it can fall into exactly the bias the paper detects.

AI safety community

Programs, schools & workshops

  • MARS V fellow · Cambridge AI Safety Hub
  • Technical AI Safety · BlueDot Impact
  • BiAlign workshop · NeurIPS
  • MLSS · Columbia University
  • ELSA workshop · privacy-preserving ML

Hands-on

I run my experiments end to end

  • Open-weight models and closed APIs
  • Fine-tuning and RL
  • Large eval runs with SLURM
  • Agentic coding, to go from idea to first results fast

Python · PyTorch · Hugging Face

Mentoring & teaching

Passing it on

  • Mentored 2 research interns · one now co-author on PrivacySkills
  • TA in probability & statistics · ENSTA Paris
  • Organised the team research seminar · PETSCRAFT

Talks & outreach

On stage, for experts and for everyone

  • TechnicalAI & Cybersecurity
  • General publicSovereignty & entrepreneurship · Cyber Founder Tour
  • OutreachInria science outreach team

Education

Mathematics at Paris-Saclay

  • MSc in Maths & AI · with CentraleSupélec & X · GPA 4.0/4.0
  • Double BSc in Maths & Physics · GPA 4.0 maths, 3.33 physics
  • Excellence scholarship · DataIA, twice

My topics

All 5 papers ↗

PrivacyMain focus

How LLMs deal with our privacy

How agents find, use and protect personal information.

Alignment

When models drift from what we want

Catching models that stray from what we intended.

Capabilities

How far a model can go

Measuring what models can really do.

Beyond the papers

Building

Doctoral entrepreneurship programme

Writing

LessWrong

Where I play with Clifford algebra as a lens for interpretability.

Latest news

2026

  1. publications

    Our paper From Observation to Intervention was accepted at the BiAlign Workshop @ NeurIPS 2026, seven weeks after the MARS kick-off.

  2. career

    I gave a talk on Jailbreaks for Uncertainty (J4U) at the AI & Cybersecurity programme.Watch ↗

  3. career

    I started the MARS V fellowship with the Cambridge AI Safety Hub.

  4. career

    I visited Trajan House in Oxford with the MARS cohort.

  5. scholar

    I attended the Machine Learning Summer School (MLSS) at Columbia University.

  6. scholar

    I completed BlueDot Impact's Technical AI Safety course.

  7. career

    I started teaching statistics tutorials (TD) at ENSTA Paris.

  8. scholar

    I followed the Inria Academy training on reinforcement learning for industry.

  9. career

    I started teaching probability tutorials (TD) at ENSTA Paris.

2025

  1. career

    I took part in a round-table on sovereignty and entrepreneurship at the Cyber Founder Tour Paris-Saclay.

  2. career

    I was a visiting researcher for several weeks at UC3M in Madrid!

  3. publications

    My first paper, on traffic forecasting, was accepted at IEEE OJITS (Q1).

  4. career

    Between March and July, I attended the ELLIS/ELSA workshop on privacy and ML, the iPoP workshop on AI development and governance, APVP 2025, the CNIL privacy research day and the INESIA scientific days.

2024

  1. career

    I joined Inria's science outreach team.

  2. career

    I started my PhD at Inria (PETSCRAFT team) on privacy in LLM agents.

  3. scholar

    I finished my MSc in Mathematics and AI with high honors.

  4. career

    The PETSCRAFT team at Inria accepted me as a PhD student on privacy and LLMs.

  5. career

    I started my AI research internship at NCHC (NARLabs, Taiwan), until September.

  6. career

    I have been accepted in the funded internship program of NARLabs (Taiwan). Thanks to the DataIA Institute for the additional grant.

2023

  1. scholar

    My MixtAI scholarship from the DataIA Institute has been renewed.

  2. career

    I started my data engineering internship at Exotrail, until August.

2022

  1. scholar

    I have received the MixtAI scholarship from the DataIA Institute.

  2. scholar

    I finished my double major licence degree in Mathematics and Physics with honors.

  3. scholar

    I was accepted into the MSc in Mathematics and AI at Paris-Saclay.