ICML 2026PastOther
ICML 2026 Workshop on Weight-Space Symmetries: from Foundations to Practical Applications
ICML 2026 Workshop WSS
- Submission deadline
- May 9, 2026, 11:59 UTCOpenReview-synced 2026-05-09 11:59 UTC (as of 2026-06-23) — extensions on OpenReview are applied automatically; verify on the website.
- Submission portal
- OpenReview
- Notes
- Topics were auto-suggested and may be imprecise — edits welcome.
Accepted papers (51)
Fetched from OpenReview (v2) on 2026-06-10.
A Geometric View of Model Merging: Quotient Fréchet Averages from Toy Models to LoRA
Access Sets Matter: Budgeting Expert Reads for Scalable Weight-Space Model Merging
Analyzing Stream Collapse in Hyper-Connections: From Diagnosis to Mitigation
Are we Merging the Right Models? Impact of Expert Training Duration on Model Merging for LLMs
Attention Weight Decomposition for Vision Model Compression
Auditing Neural Thickets with Low-Rank Routes
Beyond Pairwise: Diagnosing Higher-Order Merge Failures via Hodge Decomposition
Beyond Structural Symmetries: Linear Mode Connectivity via Neuron Identifiability
Block-Level Weight-Space Structure Persists Under Post-Training: An Empirical Study Across LLM Families
Breaking Random-Init Symmetry: Theory-Informed Initialization for ReLU Networks
Debugging ReBasin: What Limits Symmetry-Based Model Merging?
Diagonalizing the Softmax: Hadamard Initialization for Tractable Cross-Entropy Dynamics
Different Layers, Different Manifolds: Module-Wise Weight-Space Geometry in Transformer Optimization
DotResize: Reducing LLM Width via Discrete Optimal Transport-based Neuron Merging
Endpoint Symmetry for Edge Updates: Weight-Space Redundancy in GNNs on Undirected Graphs
Flow Equivariant Transformers
Generic Fibers and Functional Dimension of Multi-Head Attention
Hierarchical Mixture-of-Experts with Two-Stage Optimization
How Deep Are Deep GPs, Really? A Sharp Threshold and a Non-Gaussian Limit for Compositional GPs
How the Optimizer Shapes Learned Solutions in Equivariant Neural Networks
Iterative Magnitude Pruning Reduces Weight-Space Coupling
Low-Rank Networks Recover Weight and Functional Symmetry Better
LS-Merge: Merging Language Models in Latent Space
Meta-Merging by Checkpoint Nowcasting
Model Merging by Output-Space Projection
Model Merging via Averaged Representational Similarity
MoRE: Mixture of Reused Experts
No Global Gauge in Neural Weight Space: Branched Quotient Geometry and Atlas-Optimal Learning
Objective-Specific Privileged Bases via Full-Prefix Matryoshka Learning
On the Interplay of Priors and Overparametrization in Bayesian Neural Network Posteriors
Parameter symmetries determine representational geometry in overparameterized nonlinear networks
Pre-Normalization Momentum Governs Optimizer-Induced Rank Bias
Quantifying Symmetries: How Optimisers Impact the Functional Dimension
Rethinking the Role of Tensor Decompositions in Post-Training LLM Compression
Rotation Symmetry in Vision Quantization: The Objective Function is the Bottleneck
Scale-Equivariant Alignment: Closing the Residual Barrier After Permutation Matching
Scale-Invariant Empirical-Bayes Laplace Approximation for ReLU Networks
Sharpness-Aware Minimization Directly on the Boolean Hypercube
Shortcuts in the Tail: Debiasing via Post-Hoc Spectral Compression of Fine-Tuning Updates
SIB: Reparameterization of LLMs for Better Learning-Forgetting under SFT
Symmetry Acquisition in Predictive Coding Networks
Symmetry-Induced Non-Identifiability in Neural Circuit Inference
T-REX: Tied Recurrence Extraction
Task-Restricted Symmetries in Recurrent Weight Space
The GL(r) Gauge Symmetry of LoRA: Principal Bundle Structure, Loss Landscape Geometry, and Implications for Adapter Merging
The Role of Symmetry in Optimizing Overparameterized Networks
Toward a Type-Theoretic Framework for Linear Mode Connectivity: Univalence and Path-Finding in Weight Spaces
WARP: Weight-Space Analysis for Recovering Training Data Portfolios
Weight Space Representation Learning via Neural Field Adaptation
What Survives of Path Norms? Path-Lifting as an Intermediate Representation for ReLU Networks
WK, WV is (Linearly) All You Need: On the Necessity of the QKV Weight Triplet in Self-Attention Transformers