SafeAI@UAI 2026 — 2nd Workshop on Safe AI

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 1–22, 2026	Reviewing period
Jun 24, 2026	Author notification
Jul 9, 2026	Camera-ready deadline
Jul 16, 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, up to 4 pages (no limit on supplementary material).
Category B — Recently accepted and under-review papers at top ML / AI conferences (in the original format and length)

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. Similarly, submissions under review at other venues will be allowed, provided they do not breach any dual-submission or anonymity policies of those venues. Accepted submissions will be made publicly visible here and on OpenReview. A journal special issue may be considered.

Submission Instructions: Original submissions must comply with the UAI style requirements and use the adjusted template SafeAI format. Papers should be no longer than 4 pages, excluding references. Authors are free to include a supplementary material in the same PDF (no page limit). Papers that have already been accepted or are under-review elsewhere may be submitted using the format and page limit of the venue where they were accepted/submitted. Any supplementary material may be included in the same PDF after the references; however, reviewers may choose at their own discretion whether to consult it.

Note: Submissions are handled through OpenReview. New profiles created without an institutional email will go through a moderation process that can take up to two weeks. New profiles created with an institutional email will be activated automatically. Please plan accordingly.

Submit on OpenReview ↗

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

Tom Everitt Google DeepMind

Yarin Gal University of Oxford

Peter Mattson Google / MLCommons

Maurits Kaptein Eindhoven University of Technology

Anastasios N. Angelopoulos Arena

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.