News:

  • March 2024: I joined the Digital Humanities Lab at EPFL as a postdoc, working on the synergies between Digital Cultural Heritage and Large Language Models.
  • May 2023: I defended my Ph.D. thesis entitled "Towards Social Autotelic Artificial Agents - Formation and Exploitation of Cultural Conventions in Autonomous Embodied Artificial Agents".

Short Bio

I’m a research scientist at InstaDeep, where I develop large-scale, multi-modal foundation models—combining deep learning with structured biological data. Previously, I was a postdoctoral researcher at the Digital Humanities Lab at EPFL, exploring how Large Language Models can access and reason over historical and cultural sources not indexed on the web.

I earned my Ph.D. at the Flowers Lab (Inria), where I worked with Clément Moulin-Frier and Pierre-Yves Oudeyer on Developmental AI. My research focused on how artificial agents can self-organize cultural conventions and leverage them to autonomously generate and pursue their own goals—becoming autotelic.

Before my Ph.D., I worked in Natural Language Processing at BNP Paribas' innovation lab in New York. My academic background includes MSc degrees in Machine Learning (ISAE-Supaero) and Mechanical Engineering (EPFL).

Thesis

Project 1

Manuscript
Project 1

Defense Slides

Defense Video

Publications

Is This Collection Worth My LLM's Time? Automatically Measuring Information Potential in Text Corpora
Preprint
Preprint
T. Karch*, L. Engel*, P. Schwaller and F. Kaplan

[Paper]

Large Language Models for Interactive Exploration of Venice’s Historical Cadastre
Preprint
Preprint
T. Karch*, Y. Saydaliev*, I. di Lenardo and F. Kaplan

[Paper]

Contrastive Multimodal Learning for Emergence of Graphical Sensory-Motor Communication
Preprint
Preprint
T. Karch*, Y. Lemesle*, R. Laroche, C. Moulin-Frier and P.Y. Oudeyer

[Paper]

Vygotskian Autotelic Artificial Intelligence: Language and Culture Internalization for Human-Like AI
Nature Machine Intelligence
Nature Machine Intelligence
C.Colas*, T. Karch*, C. Moulin-Frier and P.Y. Oudeyer

[Paper][Website]

Learning to Guide and to Be Guided in the Architect-Builder Problem
ICLR 2022
ICLR2022
B. Barde*, T. Karch*, D. Nowrouzezahrai, C. Moulin-Frier, C. Pal and P.Y. Oudeyer

[Paper][Website] [Presentation]

Grounding Spatio-Temporal Language with Transformer
NeurIPS 2021
NeurIPS 2021
T. Karch*, L. Teodorescu*, K. Hoffman, C. Moulin-Frier and P.Y. Oudeyer

[Paper][Code]

Language as a Cognitive Tool to Imagine Goals in Curiosity-Driven Exploration
NeurIPS 2020
NeurIPS 2020
C. Colas*, T. Karch*, N. Lair*, J.M. Dussoux, C. Moulin-Frier, P.F. Dominey and P.Y. Oudeyer

[Paper][Code][Colab][Website]

Intrinsically Motivated Goal-Conditioned Reinforcement Learning: a Short Survey
JAIR
JAIR
C. Colas, T. Karch, O. Sigaud and P.Y. Oudeyer

[Paper]

Deep Sets for Generalization in RL
ICLR 2020 BeTRL Workshop
ICLR 2020 BeTRL Workshop
T. Karch*, C. Colas*, L. Teodorescu, C. Moulin-Frier and P.Y. Oudeyer

[Paper]

Talks

Extracting Learning Signals from Social Contexts2024, January 10
Invited talk at ISIR - Sorbonne University, 2024, January 10

[Slides]

Learning to Guide and to Be Guided in the Architect-Builder Problem2022, January 14
RL Sofa at Mila, 2022, January 14

[Video]

Vygotskian Autotelic Agents2021, July 26
Invited Talk at the Minds at Play! workshop of Cogsci 2021 , 2021, July 26

[Slides][Video]

Word Representation for Natural Language Processing (Part 2)2019, February 22
BNP Paribas AI Lunch Talk in New York, 2019, February 22

[Slides]

Word Representation for Natural Language Processing (Part 1)2019, February 22
BNP Paribas AI Lunch Talk in New York , 2019, January 28

[Slides]

Blog Posts

Guiding and Learning Without Common Ground: a New Interactive Learning Paradigm

Summary: Humans have an outstanding ability to teach and learn form each other without relying on pre-established communication protocol (not even expressing rewards). Could machine do the same? In this blog post we propose a new interactive learning paradigm to investigate this question.

Language as a Cognitive Tool: Dall-E, Humans and Vygotskian RL Agents

Summary: This blog post presents a supra-communicative view of language and advocates for the use of language as a cognitive tool to organize the cognitive development of intrinsically motivated artificial agents. We go over studies revealing the cognitive functions of language in humans, cover similar uses of language in the design of artificial agents and advocate for the pursuit of Vygotskian embodied agents - artificial agents that leverage language as a cognitive tool to structure their continuous experience, form abstract representations, reason, imagine creative goals, plan towards them and simulate future possibilities.