PhD Position F​/M Scientific Claim Verification in Research Articles

PhD Position F/M Scientific Claim Verification in Research Articles

Fonction : Doctorant

The Inria center at the University of Rennes is one of eight Inria centers and has more than thirty research teams. The Inria center is a major and recognized player in the field of digital sciences. It is at the heart of a rich ecosystem of R&D and innovation, including highly innovative SMEs, large industrial groups, competitiveness clusters, research and higher education institutions, centers of excellence, and technological research institutes.

Within the framework of a partnership (you can choose between)

• public with French National Research Agency (ANR)

Is regular travel foreseen for this post ? "travel expenses are covered within the limits of the scale in force".

Context

Large language models (LLMs) are increasingly used to assist scientific writing. While they can improve clarity and fluency, they may also introduce unsupported claims, distort results, or cite evidence in misleading ways. Detecting such problems requires more than analyzing the text alone: scientific claims must be checked against the evidence that supports them, which may appear in text, tables, figures, or cited papers.

Scientific claim verification has received growing attention in recent years. Early benchmarks such as Sci Fact introduced the task of retrieving evidence from scientific papers and determining whether it supports or refutes a claim, while later work extended this task to the more realistic and challenging multimodal setting of scientific papers, where evidence may appear in tables and figures as well as in text.

Yet, major challenges remain. Relevant evidence may be scattered across text, tables, figures, equations, code, and citations, and a correct prediction does not guarantee that a model relied on the right evidence. Developing methods that can both identify the appropriate evidence and verify claims reliably, while explaining their decisions and providing useful feedback to the user, is therefore an important step toward safer AI-assisted scientific writing.

Objective

The objective of this PhD is to develop and evaluate methods for verifying scientific claims against multimodal evidence in research articles. The work will address three main research questions, following a find > verify > explain pipeline:

• Find
  - How can we identify the evidence relevant to a scientific claim across text, tables, figures, citations, and other modalities?
• Verify
  - How can we determine whether a claim is supported by this evidence, while ensuring that the decision is grounded in the right evidence?
• Explain
  - How can verification results be explained to authors and turned into useful feedback for revising scientific claims?

The PhD is expected to produce:

• an open benchmark for multimodal scientific claim verification;
• methods for identifying and aligning claims with supporting evidence;
• evaluation protocols that assess not only prediction accuracy, but also robustness and evidence use;
• a better understanding of the strengths and limitations of models for verifying scientific content.

Technical skills and level required :

• Master's degree (or 5-year university degree) in Computer Science, Computational Linguistics, AI/NLP, or a closely related field
• Solid grounding in machine learning / deep learning; prior exposure to NLP is expected, multimodal learning (text + tables/figures) is a plus
• Proficient in Python and standard deep learning frameworks (PyTorch, Hugging Face Transformers)
• Ability to design experiments, build/annotate datasets, and critically analyze model behavior
• Comfortable reading and producing scientific writing (papers, technical reports)

Languages :

• English: fluent, spoken and written (required for scientific publishing and…