Important dates
All deadlines are end of day Anywhere on Earth (AoE).
| Apr 9, 2026 | Call for papers released |
| May 28, 2026 | Paper submission deadline |
| Jun 1–16, 2026 | Reviewing period |
| Jun 18, 2026 | Author notification |
| Jul 2, 2026 | Camera-ready deadline |
| Jul 9, 2026 | Program & papers online |
| Aug 21, 2026 | Workshop day — Amsterdam |
Call for papers
We invite two categories of submissions:
- Category A — Original papers: UAI-style formatting, same length restrictions as the main conference.
- Category B — Extended abstracts: 2 pages, for papers accepted elsewhere and for invited presentation pitches.
Review process: Single-blind peer review by the program committee. At least one author must attend in person. There will be no proceedings, so authors are free to submit their work elsewhere. A journal special issue may be considered.
Topics of interest
We solicit submissions presenting original research, theoretical results, and applied work within the following topics (list not exhaustive).
Foundations for safe and agentic AI
- Foundations of Safe AI across learning paradigms, including reinforcement and continual learning
- Sequential decision-making under uncertainty, including Bayesian decision theory and POMDPs
- Objective specification, reward design, and alignment for agentic systems
- Test-time scaling, adaptive inference, and safety implications of increasingly capable agents
Uncertainty, robustness, and control
- Uncertainty quantification, calibration, and propagation in decision-making
- Risk-sensitive planning and long-horizon safety under uncertainty
- Robustness to distribution shift, adversarial conditions, and non-stationary environments
- Safety and control in RL-based agents, including safe exploration and learning-based control
Interpretability, auditing, and evaluation
- Interpretability and explainability of agent policies, representations, and behaviors
- Methods to audit, measure, monitor, and evaluate agentic AI systems
- Benchmarks, metrics, and case studies for safe and aligned agents
- Adversarial evaluation and stress testing of agentic systems, including red teaming
Engineering, deployment, and applications
- Agent failure modes: reward hacking, goal misgeneralization, agent hijacking, and emergent behaviors
- Human-in-the-loop and oversight mechanisms for agentic systems
- Deployment-time safety of AI in safety-critical and autonomous systems
Invited speakers
More speakers to be announced.
Panel discussion
In-person speakers and leading researchers from academia and industry will discuss:
- Uncertainty and decision-making in autonomous agents
- Interpretability and oversight for action-taking systems
- Alignment and objective specification over long horizons
- Scalable evaluation and control of agentic systems
Motivation
Ensuring the safety of AI systems has become a foundational challenge as they are increasingly deployed in high-stakes domains and subject to regulatory and societal expectations. Beyond accuracy and performance, Safe AI concerns reliability, robustness, interpretability, and alignment across the full system lifecycle.
This workshop focuses on technical and socio-technical challenges in Safe AI, with particular attention to agentic AI systems — learning-based systems that autonomously select actions, interact with an environment, and pursue objectives over time. In such systems, actions influence data, feedback, and behavior over extended horizons, making safety tightly linked to sequential decision-making, adaptation, and interaction. These challenges place uncertainty, robustness, and interpretability at the center of agentic safety, directly connecting them to foundational questions studied by the UAI community.
Organizers
