Call for Papers

Submission Portal	OpenReview
Deadline	April 30, 2026 (AOE) → May 8, 2026 (AOE)
Author notification	May 22, 2026, AOE

We welcome all original research papers of up to 4 pages in length (excluding references and supplementary material), with one additional page allowed for the camera-ready version. We also permit papers that have been recently published or are under submission to another venue.

All accepted papers will be presented as posters; a subset of especially noteworthy papers will be selected for oral presentations. We will present a Best Paper Award to recognize outstanding research.

This workshop is non-archival; even though all accepted papers will be available on OpenReview, there are no formally-published proceedings.

Submission Instructions

• Submissions should be a single file in .pdf format, submitted via OpenReview (link TBA).
• We only accept submissions prepared using LaTeX. Use the official ICML 2026 style files with icml2026_weightsymmetry.sty in place of the standard ICML style file (icml2026.sty).
• The review process is double-blind; ensure all submissions are properly anonymised.
• All submissions will receive at least three reviews. Conflicts of interest must be disclosed.
• Supplementary material is permitted but reviewers are not required to read it – ensure your submission is self-contained.

Scope and topics

We aim to connect theoretical and practical research on weight-space symmetries and welcome contributions from across this spectrum. We are interested in work that advances the theoretical understanding of weight-space structure, empirical studies that shed light on symmetry effects, practical methods that exploit symmetries, and approaches that make these methods work for modern models at scale.

We welcome submissions on topics including, but not limited to:

Characterizing Weight-Space Symmetries

• Hidden and undiscovered symmetries in neural network parameter spaces.
• Symmetries in modern architectures (MoE, SSMs, LoRA, multi-modal networks).
• Beyond exact symmetries: approximate, data-dependent, and distributional symmetries.
• Connections between weight-space symmetries and representation-space symmetries.

Loss Landscape and Optimization

• Geometry and topology of the loss landscape induced by weight-space symmetries.
• Effect of weight-space symmetries on optimization trajectories and convergence.
• Scalable symmetry-aware and symmetry-breaking optimization: leveraging symmetries for better initialization, training acceleration, and trajectory reuse.
• Implicit bias, generalization, and expressivity through the lens of weight-space symmetry.

Linear Mode Connectivity and Model Merging

• Permutation- and symmetry-based alignment methods for model merging.
• Extensions of linear mode connectivity to modern large-scale architectures.
• Merging models trained from different initializations, tasks, or modalities at scale.
• Understanding of when and why merging succeeds or fails.

Weight-Space Learning

• Symmetry-aware analysis and comparison of individual models and model populations.
• Equivariant and invariant architectures for processing neural network weights, and weight augmentation methods.
• Generative models over weight spaces: sampling, synthesis, and interpolation.
• Applications to accelerating training, model merging, model editing, and other emerging cases.

Other Applications of Weight-Space Symmetries

• Model compression, quantization, and efficient architectures.
• Uncertainty quantification and Bayesian inference.
• Model safety, interpretability, and mechanistic analysis.
• Any other setting where weight-space symmetries provide theoretical or practical insights.