Overview
Visit the inaugural ICLR 2025 workshop website.
| Venue | NeurIPS 2026 |
| Date | To be announced |
| Location | Paris and Remote |
| Submissions | Extended abstracts and full papers via OpenReview; deadline: September 1, 2026 (Anywhere on Earth) |
Machine learning has revolutionized how we learn from scientific data, yet it has rarely turned that same population-level lens on its own products. This workshop aims to close that gap by treating neural network artifacts as a data modality in their own right.
Today’s model repositories contain immense distributed knowledge encoded not only in neural network weights, but also in gradients, intermediate representations, optimization trajectories, and other computational traces. We refer to these collectively as neural artifacts. Learning from populations of these artifacts can help us compare, search, explain, modify, control, and synthesize models.
Following the inaugural ICLR 2025 workshop, this second edition broadens the scope beyond weights and places greater emphasis on standardized datasets, benchmarks, tasks, neural lineages, and AI supply chains. Our goal is to connect communities working on model merging, meta-learning, mechanistic interpretability, neural architecture search, neural fields, and related areas under a shared data-centric perspective.
Workshop Themes
- Datasets and benchmarks: Standardized model zoos, evaluation protocols, and new tasks for learning from neural artifacts.
- Foundations and theory: Structure, symmetries, scaling laws, and specialized architectures such as equivariant metanetworks.
- Model analysis and dynamics: Inferring behavior, generalization, safety, robustness, fairness, memorization, and backdoors from artifacts; understanding learning dynamics and interpretability.
- Model synthesis and control: Generating and editing models through hypernetworks, task arithmetic, merging, steering, and related techniques.
- Model search and selection: Navigating model populations to select models for inference, fine-tuning, or transfer without expensive retraining or evaluation.
- Model populations and AI supply chains: Mapping model lineages, trends, knowledge gaps, and supply-chain effects through tools such as model atlases.
Research Goals and Key Questions
This workshop will explore questions such as:
- How should neural artifacts be represented, compared, and modeled across architectures and training runs?
- What can weights and computational traces reveal about model behavior, provenance, safety, and learning dynamics?
- How can populations of models support efficient search, selection, transfer, merging, editing, and generation?
- Which datasets, benchmarks, and evaluation protocols are needed to make progress measurable and reproducible?
- How can insights from theory, interpretability, neural fields, and AI supply chains strengthen one another?
See the Call for Papers for submission tracks and the full list of topics.
