Organizers
Overview
Large Vision-Language Models (LVLMs) and Large Language Models (LLMs) are increasingly used for visual question answering, retrieval, code generation, mathematical reasoning, decision support, and other high-impact applications. Despite their impressive capabilities, these models can generate plausible but incorrect information, remain overconfident when they fail, and be manipulated through adversarial prompts, poisoned data, jailbreaks, or attacks on their multimodal inputs.
This tutorial presents a unified perspective on trustworthy LLMs and LVLMs. We connect three complementary questions: How can failures and hallucinations be detected? How can their uncertainty and predictive risk be quantified? and How can models and inference pipelines be defended against adversarial manipulation? The tutorial is designed to be accessible to newcomers while also presenting recent research directions, practical evaluation tools, and open problems.
Learning objectives
- Understand hallucinations: distinguish intrinsic, extrinsic, visual, factual, and reasoning-related failures.
- Quantify uncertainty: understand why token probability alone is insufficient and compare token-, sequence-, and semantic-level uncertainty.
- Model adversarial threats: identify vulnerabilities in prompts, visual inputs, context windows, retrieval pipelines, and fine-tuning data.
- Evaluate defenses: study grounding, verification, abstention, robust alignment, adversarial training, purification, and safety-aware decoding.
- Apply the ideas: use practical libraries and notebooks to audit model safety and evaluate reliability.
Program
- 9:00 - 9:15: Introduction: Towards Trustworthy LLMs and LVLMs — Andrei Bursuc
- 9:15 - 9:30: Sources of Uncertainty and Their Evolution in Modern AI Models — Andrei Bursuc
- 9:30 - 9:40: Understanding Hallucinations in LLMs and LVLMs — Gianni Franchi
- 9:40 - 10:10: Security, Attacks, and Threats in LLMs and LVLMs — Mael Jenny
- 10:10 - 10:30: ☕ Coffee Break ☕
- 10:30 - 11:50: Uncertainty Quantification Techniques for Large Language Models — Gianni Franchi
- 11:50 - 12:00: Conclusions, Open Questions, and Q&A — All organizers
Outline
Introduction: Towards Trustworthy LLMs and LVLMs
We begin by introducing the main reliability challenges of modern Large Language Models (LLMs) and Large Vision-Language Models (LVLMs). We discuss why trustworthy deployment requires understanding not only model performance, but also uncertainty, hallucinations, and security.
Part I: Sources of Uncertainty and Their Evolution
We introduce the different sources of uncertainty that arise in machine learning systems and discuss how uncertainty evolves when moving from classical predictive models to modern foundation models. This provides the conceptual foundations for understanding reliability in LLMs and LVLMs.
Part II: Hallucinations in LLMs and LVLMs
We introduce hallucinations as a central failure mode of generative models and discuss their relationship with uncertainty. We highlight why a model can generate plausible and confident answers while still being incorrect, and discuss the challenges involved in detecting such failures.
Part III: Security, Attacks, and Threats
We present the main security challenges affecting LLMs and LVLMs, including adversarial attacks, prompt injection, jailbreaks, and other vulnerabilities. We discuss how these threats affect the reliability and safe deployment of foundation models.
Part IV: Uncertainty Quantification Techniques for Large Language Models
This part provides an in-depth overview of uncertainty quantification techniques for LLMs. We discuss token- and sequence-level confidence, calibration, sampling-based approaches, semantic uncertainty, ensembles, Bayesian methods, and uncertainty-aware abstention. We emphasize the strengths and limitations of the different approaches and their use for detecting unreliable predictions.
Conclusions and Open Questions
We conclude by connecting uncertainty, hallucinations, and security, and discuss open research questions for building more reliable and trustworthy LLMs and LVLMs.
Intended audience and prerequisites
The tutorial targets academic researchers, graduate students, machine learning engineers, and industry practitioners interested in AI safety, security, trustworthy machine learning, and multimodal foundation models. It is particularly relevant to researchers and engineers deploying generative models in high-stakes environments.
Recommended prerequisites:
- Graduate-level knowledge of deep learning and optimization.
- Basic familiarity with Transformers, attention, and autoregressive generation.
- Basic Python knowledge and familiarity with the Hugging Face ecosystem for the practical session.
Relation to prior tutorials and short courses
Previous tutorials have studied individual dimensions of trustworthy generative AI. Some have focused on uncertainty quantification and reliability for deep learning, while others have addressed hallucinations, factuality, jailbreaks, or agentic attacks. Examples include The Many Faces of Reliability of Deep Learning for Real-World Deployment, the AAAI tutorial on uncertainty quantification for large language models, and the Jailbreaking LLMs and Agentic Systems tutorial.
Our tutorial differs by placing hallucination analysis, semantic uncertainty, adversarial threat modeling, and robust defense within one unified framework. It also emphasizes practical evaluation: participants will learn not only how failures arise, but also how to measure predictive risk and test whether a defense remains effective against adaptive attacks.
Selected References
- Laurent, O., Lafage, A., Tartaglione, E., Daniel, G., Martinez, J. M., Bursuc, A., & Franchi, G. Packed-Ensembles for Efficient Uncertainty Estimation. ICLR, 2023.
- Lafage, A., Laurent, O., Gabtni, F., & Franchi, G. Torch-Uncertainty: A Deep Learning Framework for Uncertainty Quantification. NeurIPS, 2025.
- Franchi, G., Laurent, O., Leguéry, M., Bursuc, A., Pilzer, A., & Yao, A. Make Me a BNN: A Simple Strategy for Estimating Bayesian Uncertainty from Pre-trained Models. CVPR, 2024.
- Hoche, J., Bursuc, A., Brellmann, D., Louppe, G., Izmailov, P., Yao, A., & Franchi, G. Improving Semantic Uncertainty Quantification in LVLMs with Semantic Gaussian Processes. arXiv, 2025.
- Xiao, J., Yao, A., Li, Y., & Chua, T. S. Can I Trust Your Answer? Visually Grounded Video Question Answering. CVPR, 2024.
- Zou, A., Wang, Z., Kolter, J. Z., & Fredrikson, M. Universal and Transferable Adversarial Attacks on Aligned Language Models. 2023.
- Ji, Z., Lee, N., Frieske, R., et al. Survey of Hallucination in Natural Language Generation. ACM Computing Surveys, 2023.
- Kuhn, L., Gal, Y., & Farquhar, S. Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation in Natural Language Generation. ICLR, 2023.
A tutorial connecting academia and industry
Organized by researchers from AMIAD, ENSTA Paris, Safran Tech, Inria, valeo.ai, and the National University of Singapore.