3IA PhD/Postdoc Seminar #56

Flash presentation | AI-based precision oncology to monitor response of metastatic cancer to immunotherapy using PET/CT imaging

Abstract: Can LLMs provide reliable prior knowledge for Bayesian inference? While LLM-based prior elicitation offers a scalable alternative to expert-driven approaches, incorrect or hallucinated information can negatively impact posterior inference. We investigate these challenges in the context of metastatic lung cancer patient data provided by the Centre Antoine Lacassagne (CAL). Our goal is to develop a principled framework for trustworthy knowledge-driven Bayesian inference with LLMs, enabling informative prior construction while ensuring robustness and reliability.

Solving the Traveling Salesman Problem with Positional Encoding

Abstract: Recent advances in Neural Combinatorial Optimization for the Traveling Salesman Problem (TSP) rely heavily on coordinate-based feature projections. In this work, we propose Positional Encoding-based Neural Solvers (PENS), a class of transformer-based solvers that treat spatial coordinates as positional encodings rather than input features. By adapting ALiBi and RoPE, modern positional encodings originally developed for large language models, our framework inherits powerful invariances and locality biases. To maintain attention resolution in large instances, we introduce a simple yet effective rescaling heuristic that further boosts performance. Trained only on TSP-100, PENS achieves state-of-the-art results for instances with up to 10 000 cities, a scale that was previously dominated by methods requiring graph sparsification. These findings demonstrate that positional encodings provide effective inductive biases for neural combinatorial optimization.

Rethinking Link Prediction in Knowledge Graphs: Instability, Semantic Awareness, and Neuro-Symbolic Approaches

Abstract: Knowledge Graph Embedding Models (KGEMs) have become the dominant approach for link prediction (LP) in knowledge graphs. However, this task is still mostly evaluated through rank-based metrics such as MRR or Hits@K, leaving open important questions regarding the operationalization of LP in real-world settings. In this talk, I revisit LP through two complementary perspectives: semantic validity and prediction stability. First, I show how semantic-aware evaluation and training strategies, leveraging relation signatures such as domains and ranges, provide a more comprehensive understanding of model behavior and improve both predictive quality and semantic correctness. Second, I highlight an important but underexplored issue: the instability of KGEMs across random seeds and training conditions, raising concerns about the reliability and reproducibility of LP approaches in real-world settings. Altogether, this talk advocates for a rethinking of link prediction evaluation and training, towards more reliable, semantically grounded, and ultimately more neuro-symbolic approaches to knowledge graph completion.

Open discussion about all the contributions