Research Groups: Data, Intelligence, and Graph Team, Télécom Paris,
France. BNP Paribas, France

Advisors: Mehwish Alam, Bérénice Jaulmes, Jean-Christophe Arouette.
Scientific Context.

Recent progress in Large Language Models (LLMs) has led to remarkable
advances in Chain-of-Thought (CoT) reasoning, the step-by-step
generation of intermediate thoughts to reach a final answer. However,
the reliability and interpretability of these reasoning chains remains
an open challenge. This internship will contribute to the growing
scientific effort to verify and evaluate CoT reasoning and its
verifiers through the analysis and development of benchmark datasets
and evaluation metrics. The goal is to strengthen our understanding of
how LLMs reason, identify where they fail, and provide a basis for
designing methods to measure reasoning quality more accurately.

The project will explore existing benchmarks, which provide large-scale
annotated datasets for reasoning verification across domains like
mathematics, physics, and commonsense reasoning. Each of these
benchmarks introduces different verification methodologies; for
instance, PRM800k uses fine-grained human annotations for every
reasoning step, while THINK-Bench introduces precision and recall
metrics for key logical steps. The intern will analyze the advantages
and drawbacks of these datasets, such as scalability, annotation
reliability, and domain coverage, and investigate how they can be
extended or combined for more comprehensive reasoning evaluation.


Candidate Profile.

-   Currently pursuing M2 in the field of Artificial
    Intelligence/Machine Learning
-   Good programming skills, such as in Python (incl. Pytorch).
-   Knowledge of Large Language Models is a plus but not required;
    however, interest in learning and keeping themselves up-to-date
    with upcoming trends in the field is required.
-   Good communication skills, especially in English.

Required Documents.

-   A full CV
-   A motivation letter expressing your interest in the position and
    relevant experience
-   A transcript of records


Contacts.

Please send the complete required documents to

Mehwish Alam (mehwish.alam@telecom-paris.fr),
Bérénice Jaulmes (berenice.jaulmes@ip-paris.fr), and
Jean-Christophe Arouete jean-christophe.arouete@bnpparibas.com)

in an email with the subject starting with "[M2Internship-CoT]".