Continual Adaptation at Scale — ICML 2026 Workshop

About

Training foundation models is currently so costly that only few can afford it. The immense data, compute, and energy demands are increasingly unsustainable. Continual adaptation offers a viable alternative, where AI models can learn quickly and continually through every day interactions, just like humans and animals.

Unfortunately, foundation models lack this rapid adaptability: new behavior can be induced by prompting or fine-tuning, but there are no easy ways to quickly shape the behavior — for instance, to permanently add, remove, or modify their skill set in a sustainable way. This workshop aims to discuss new research directions that will enable fast continual adaptation at scale to drive more sustainable AI.

Key Research Directions

Scale & Efficiency: The immense scale of foundation models makes traditional continual learning methods infeasible. This reshapes core questions about where plasticity should reside: must we update weights, modify architectures, or rely purely on in-context learning—and which approaches, alone or combined, are sufficient?
Lifecycle & Alignment: Foundation models undergo a hierarchical lifecycle of pre-training, instruction tuning, and alignment. We should determine if alignment and knowledge can be disentangled, or if the alignment tax is an inevitable cost of continual adaptation.
Multimodality: Foundation models are inherently multimodal, introducing additional challenges like cross-modal drift, where updating vision encoders can degrade language reasoning which can then require a rigorous re-evaluation of how we measure forward and backward transfer.
Deliberate Forgetting: While traditional continual learning focuses on knowledge accumulation, foundation models also require selective data removal for privacy, safety, or updating facts. This necessitates a new framework that treats intentional forgetting as a primary objective rather than merely avoiding catastrophic interference.

Invited Speakers

Razvan Pascanu

Google DeepMind / MILA

Sara Hooker

Adaption

Bing Liu

University of Illinois at Chicago

Stephanie Chan

Google DeepMind

Colin Raffel

University of Toronto / Vector Institute / Hugging Face

Jaehong Yoon

Nanyang Technological University

Schedule

All times are local to the venue.

Time	Type	Activity
08:00 – 08:10	Opening	Opening Remarks
08:10 – 08:50	Invited Talk	Razvan Pascanu
08:50 – 09:35	Oral Session	Oral Session 1
09:35 – 10:15	Invited Talk	Colin Raffel
10:15 – 11:00	Break	Poster Session 1 & Coffee
11:00 – 11:40	Invited Talk	Sara Hooker
11:40 – 12:20	Invited Talk	Jaehong Yoon
12:20 – 13:30	Break	Lunch Break
13:30 – 14:10	Invited Talk	Bing Liu
14:10 – 14:55	Oral Session	Oral Session 2
14:55 – 15:40	Break	Poster Session 2 & Coffee
15:40 – 16:20	Invited Talk	Stephanie Chan
16:20 – 16:30	Awards	Best Paper Awards
16:30 – 17:00	Panel	Speaker Panel

Call for Papers

We invite submissions on the following topics:

Frameworks for Adaptation: Reframing Post-Training and Alignment (SFT/RLHF) as continual adaptation to mitigate vulnerabilities from distribution shifts (e.g., alignment drift), and tailoring strategies to each lifecycle phase.
Mechanisms of Adaptation: Analyzing parametric (PEFT), non-parametric (RAG), and semi-parametric (ICL) methods to design modular architectures optimized for low-cost adaptation.
Theoretical Foundations of Adaptation: Understanding the fine-tuning optimization landscape and how concepts are stored and modified in high-dimensional space.
Evaluation & Benchmarks for Adaptation: Moving beyond Class-Incremental Learning to realistic scenarios (temporal news shifts, evolving code, agent workflows) and establishing metrics for efficiency, alignment retention, and targeted removal.

Important Dates

Submission Deadline: April 30, 2026 (Anywhere on Earth)
Review Period: May 1-12, 2026
Notification of Acceptance: May 13, 2026
Camera-Ready Deadline: June 15, 2026
Workshop Date: TBA

Submission Guidelines

Submitted papers are composed of a main body, which can be up to 4 pages long, followed by unlimited pages for references and an appendix, all in a single file. For details concerning the format of the papers, please see the LaTeX style files (link coming soon). All submissions must be via OpenReview (link coming soon) and anonymized, otherwise, they will automatically be rejected. In particular, any submission whose main body goes over the 4 page limit will be automatically rejected.