Hello,
I am seeking a candidate for a fully funded PhD position focused on fast and efficient methods for modern natural language processing. A full description of the project is provided below.Best,
Caio
PhD Thesis: Toward Fast and Efficient Methods for Modern
Natural Language Processing
- Position location: Rennes, France
- Lab: IRISA, team Archimedia
- Supervisors: Caio Corro, Guillaume Gravier
- Funding is already secured via the French National
Research Agency
Large language models (LLMs) and deep contextual embeddings (i.e.
BERT and its variants) are ubiquitous in natural language
processing (NLP). Unfortunately, they come at a cost : both
fine-tuning and decoding are costly procedures, especially when
one does not have access to sufficient GPU resources.
The aim of this PhD project is to propose novel methods that
allow to build faster models, or to adapt models at test-time
without costly fine-tuning procedures. To this end, we will
take inspiration in previous works, for example, but not limited
to :
- Speculative decoding, that allows to leverage GPU
parallelization capabilities to speed up text generation speed
[1,2,3];
- Direct preference optimization, that allows to solve the
standard reinforcement learning from human feedback (RLHF) problem
with a simpler (and faster in practice) training objective [4];
- Test-time alignment, that relies on a small model to guide a
large one [5];
- Approximate inference algorithms that can fully leverage GPU
parallelization capabilities [6,7];
- Few-shot adaptation methods [8];
- etc etc.
The specific applications and targeted methods will be decided
jointly with the candidate. The PhD is fully funded by the
French National Research Agency via the SemiAmor research grant.
The candidate must have an interest for deep learning and natural
language processing, but also for algorithmic and optimization.
The candidate is expected to have either a strong computer science
background with an interest for applied mathematics, or a strong
mathematical background with some knowledge in deep learning and
Python+Pytorch. As the goal of the project is to propose novel
methods, the candidate is expected to be able to develop his own
code and to hack libraries like Pytorch and HuggingFace, i.e. the
project will require to go beyond a simple use of tools.
Outcomes of this project are expected to be published in the main
natural language processing conferences/journals (*ACL/EMNLP/TACL)
and/or main machine learning conferences/journals
(NeurIPS/ICLR/ICML/AISTATS/TMLR).
Supervisors:
- Caio Corro (INSA Rennes, IRISA) https://caio-corro.fr/
- Guillaume Gravier (CNRS, IRISA)
To apply, please send an email at caio....@irisa.fr with the
following documents :
- a CV
- last year of bachelor (licence) and master grades
- a short description of your main interests in NLP and computer
science
If you have any question, feel free to send an email to
caio....@irisa.fr
[1] Fast Inference from Transformers via Speculative Decoding
(Yaniv Leviathan, Matan Kalman, Yossi Matias)
https://proceedings.mlr.press/v202/leviathan23a.html
[2] QSpec: Speculative Decoding with Complementary Quantization
Schemes (Juntao Zhao, Wenhao Lu, Sheng Wang, Lingpeng Kong, Chuan
Wu) https://aclanthology.org/2025.emnlp-main.240/
[3] Cactus: Accelerating Auto-Regressive Decoding with Constrained
Acceptance Speculative Sampling (Yongchang Hao, Lili Mou)
https://openreview.net/forum?id=lpUIkCAy9p
[4] Direct Preference Optimization: Your Language Model is
Secretly a Reward Model (Rafael Rafailov, Archit Sharma, Eric
Mitchell, Stefano Ermon, Christopher D. Manning, Chelsea Finn)
https://papers.nips.cc/paper_files/paper/2023/hash/a85b405ed65c6477a4fe8302b5e06ce7-Abstract-Conference.html
[5] KAD: A Framework for Proxy-based Test-time Alignment with
Knapsack Approximation Deferral (Ayoub Hammal, Pierre Zweigenbaum,
Caio Corro) https://aclanthology.org/2026.eacl-long.179/
[6] Sinkhorn Distances: Lightspeed Computation of Optimal
Transport (Marco Cuturi)
https://papers.nips.cc/paper_files/paper/2013/hash/af21d0c97db2e27e13572cbf59eb343d-Abstract.html
[7] Bregman Conditional Random Fields: Sequence Labeling with
Parallelizable Inference Algorithms (Caio Corro, Mathieu Lacroix,
Joseph Le Roux) https://aclanthology.org/2025.acl-long.1430/
[8] Few-shot domain adaptation for named-entity recognition via
joint constrained k-means and subspace selection (Ayoub Hammal,
Benno Uthayasooriyar, Caio Corro)
https://aclanthology.org/2025.coling-main.662/