As the landscape of artificial intelligence evolves, ensuring the safety and alignment of superintelligent language models (LLMs) is paramount. This workshop will delve into the theoretical foundations of LLM safety. This could include topics like the Bayesian view of LLM safety versus the RL view of safety and other theories.
The flavor of this workshop is futuristic, focusing on how to ensure a superintelligent LLM/AI remains safe and aligned with humans. This workshop is a joint effort of the Simons Institute and IVADO.
Key Topics:
- Bayesian Approaches to LLM Safety
- Reinforcement Learning Perspectives on Safety
- Theoretical Frameworks for Ensuring AI Alignment
- Case Studies and Practical Implications
- Future Directions in LLM Safety Research
This workshop is part of the programming for the thematic semester on Large Language Models and Transformers, organized in collaboration with the Simons Institute for the Theory of Computing.
Travel grants are available to attend the event in California.
Workshops will also be available online and live (registration required).
Organizers
AGENDA
Monday, Apr. 14th, 2025
9 – 9:15 a.m.: Coffee and Check-In
9:15 – 9:30 a.m.: Welcome Address
9:30 – 10:30 a.m.: Simulating Counterfactual Training
Roger Grosse (University of Toronto)
10:30 – 11 a.m.: Break
11 a.m. – 12 p.m.: AI Safety via Inference-Time Compute
Boaz Barak (Harvard University and OpenAI)
12 – 2 p.m.: Lunch (on your own)
2 – 3 p.m.: Future Directions in AI Safety Research
Dawn Song (UC Berkeley)
3 – 3:30 p.m.: Break
3:30 – 4:30 p.m.: Scalable AI Safety via Efficient Debate Games
Georgios Piliouras (Singapore University of Technology and Design)
4:30 – 6 p.m.: Reception
Tuesday, Apr. 15th, 2025
9:30 – 10 a.m.: Coffee and Check-In
10 – 11 a.m: Full-Stack Alignment
Ryan Lowe (Meaning Alignment Institute)
11 – 11:30 a.m.: Break
11:30 a.m. – 12:30 p.m.: Can We Get Asymptotic Safety Guarantees Based On Scalable Oversight?
Geoffrey Irving (UK AI Security Institute)
12:30 – 2:30 p.m.: Lunch (on your own)
2:30 – 3:30 p.m.: Amortised Inference Meets LLMs: Algorithms and Implications for Faithful Knowledge Extraction
Nikolay Malkin (University of Edinburgh)
3:30 – 4 p.m.: Break
4 – 5 p.m.: Special Lecture – Richard M. Karp Distinguished Lecture
Yoshua Bengio (IVADO – Université de Montréal – Mila)
5 – 6 p.m.: Panel Discussion
Yoshua Bengio (IVADO – Mila – Université de Montréal), Dawn Song (UC Berkeley), Roger Grosse, Geoffrey Irving, Siva Reddy (IVADO – Mila – McGill University)
Wednesday, Apr. 16th, 2025
8:30 – 9 a.m.: Coffee and Check-In
9 – 10 a.m.: Talk By
Siva Reddy (IVADO – McGill University – Mila)
10 – 10:15 a.m.: Break
10:15 – 11:15 a.m.: Adversarial Training for LLMs’ Safety Robustness
Gauthier Gidel (IVADO – Université de Montréal – Mila)
11:15 – 11:30 a.m.: Break
11:30am – 12:30 p.m.: Talk By
Zico Kolter (Carnegie Mellon University)
12:30 – 2 p.m.: Lunch (on your own)
2 – 3 p.m.: Causal Representation Learning: A Natural Fit for Mechanistic Interpretability
Dhanya Sridhar (IVADO – Université de Montréal – Mila)
3 – 3:15 p.m.: Break
3:15 – 4:15 p.m.: Out of Distribution, Out of Control? Understanding Safety Challenges in AI
Aditi Raghunathan (Carnegie Mellon University)
Thursday, Apr. 17th, 2025
9 – 9:30 a.m.: Coffee and Check-In
9:30 – 10:30 a.m.: LLM Negotiations and Social Dilemmas
Aaron Courville (IVADO – Université de Montréal – Mila)
10:30 – 11 a.m.: Break
11 a.m. – 12 p.m: Scalably Understanding AI With AI
Jacob Steinhardt (UC Berkeley)
12 – 1:45 p.m.: Lunch (on your own)
1:45 – 2:45 p.m.: Controlling Untrusted AIs with Monitors
Ethan Perez (Anthropic)
2:45 – 3 p.m.: Break
3 – 4 p.m.: What Can Theory of Cryptography Tell us About AI Safety
Shafi Goldwasser (UC Berkeley)
4 – 5 p.m.: Assessing the Risk of Advanced Reinforcement Learning Agents Causing Human Extinction
Michael Cohen (UC Berkeley)
Friday, Apr. 18th, 2025
8:30 – 9 a.m.: Coffee and Check-In
9 – 10 a.m.: Safeguarded AI Workflows
David Dalrymple (Advanced Research + Invention Agency)
10 – 10:15 a.m.: Break
10:15 – 11:15 a.m.: AI Safety: LLMs, Facts, Lies, and Agents in the Real World
Chris Pal (IVADO + Polytechnique + Mila + UdeM DIRO + CIFAR + ServiceNow)
11:15 – 11:30 a.m.: Break
11:30 a.m. – 12:30 p.m.: Measurements for Capabilities and Hazards
Dan Hendrycks (Center for AI Safety)
12:30 – 2 p.m.: Lunch (on your own)
2 – 3 p.m.: Theoretical and Empirical aspects of Singular Learning Theory for AI Alignment
Daniel Murfet (Timaeus)
3 – 3:30 p.m.: Break
3:30 – 4:30 p.m.: Probabilistic Safety Guarantees Using Model Internals
Jacob Hilton (Alignment Research Center)
4:30 – 4:45 p.m: Closing Remarks
