We analyze a learning-to-optimize (L2O) algorithm that exploits parameter space symmetry to enhance optimization efficiency. Prior work has shown that jointly learning symmetry transformations and local updates improves meta-optimizer performance. Supporting this, our theoretical analysis demonstrates that even without identifying the optimal group element, the method locally resembles Newton's method. We further provide an example where the algorithm provably learns the correct symmetry transformation during training. To empirically evaluate L2O with teleportation, we introduce a benchmark, analyze its success and failure cases, and show that enhancements like momentum further improve performance. Our results highlight the potential of leveraging neural network parameter space symmetry to advance meta-optimization.

Foundational language models show a remarkable ability to learn new concepts during inference via context data. However, similar work for images lag behind. To address this challenge, we introduce FLoWN, a flow matching model that learns to generate neural network parameters for different tasks. Experiments verify that FLoWN attains various desiderata for a meta-learning model. In addition, it matches or exceeds baselines on in-distribution tasks, provides better initializations for classifier training, and is performant on out-of-distribution few-shot tasks while having a fine-tuning mechanism to improve performance. However, its application to neural network weights has not been explored. By leveraging the principled, yet versatile training of FM, we aim to generate task-specific weights on novel tasks. Multiple approaches have been tried to generate weights capable of few-shot learning (FSL), motivated by its speed compared to conventional training. For instance, various diffusion-based approaches (Soro et al., 2024; Zhang et al., 2024; Wang et al., 2024) have been used to generate neural network weights. However, flexibility is limited by its restriction to Gaussian processes and a sluggish inference speed.

The availability of large, structured populations of neural networks - called 'model zoos' - has led to the development of a multitude of downstream tasks ranging from model analysis, to representation learning on model weights or generative modeling of neural network parameters. However, existing model zoos are limited in size and architecture and neglect the transformer, which is among the currently most successful neural network architectures. We address this gap by introducing the first model zoo of vision transformers (ViT). To better represent recent training approaches, we develop a new blueprint for model zoo generation that encompasses both pre-training and fine-tuning steps, and publish 250 unique models. They are carefully generated with a large span of generating factors, and their diversity is validated using a thorough choice of weight-space and behavioral metrics. To further motivate the utility of our proposed dataset, we suggest multiple possible applications grounded in both extensive exploratory experiments and a number of examples from the existing literature. By extending previous lines of similar work, our model zoo allows researchers to push their model population-based methods from the small model regime to state-of-the-art architectures. We make our model zoo available at github.com/ModelZoos/ViTModelZoo.

Re-using trained neural network models is a common strategy to reduce training cost and transfer knowledge. Weight space learning - using the weights of trained models as data modality - is a promising new field to re-use populations of pre-trained models for future tasks. Approaches in this field have demonstrated high performance both on model analysis and weight generation tasks. However, until now their learning setup requires homogeneous model zoos where all models share the same exact architecture, limiting their capability to generalize beyond the population of models they saw during training. In this work, we remove this constraint and propose a modification to a common weight space learning method to accommodate training on heterogeneous populations of models. We further investigate the resulting impact of model diversity on generating unseen neural network model weights for zero-shot knowledge transfer. Our extensive experimental evaluation shows that including models with varying underlying image datasets has a high impact on performance and generalization, for both in-and out-of-distribution settings. When training neural networks for computer vision applications, we follow a dominant paradigm of pre-training and fine-tuning (Pan & Y ang, 2010; Y osinski et al., 2014), either by using pre-trained models trained from single datasets (Mensink et al., 2021) or pre-trained foundation models, which can be used for fine-tuning to multiple downstream tasks (Bommasani et al., 2021; Qiu et al., 2024). Given the vast amounts of pre-trained models, which have been deployed and released publicly on platforms such as Pytorch Hub or Huggingface, the research community has extended this paradigm by proposing the transfer or distillation of knowledge not only from one model but rather from a collection or population of pre-trained models.

Model merging acquires general capabilities without extra data or training by combining multiple models' parameters. Previous approaches achieve linear mode connectivity by aligning parameters into the same loss basin using permutation invariance. In this paper, we introduce Model Assembly Learning (MAL), a novel paradigm for model merging that iteratively integrates parameters from diverse models in an open-ended model zoo to enhance the base model's capabilities. Unlike previous works that require identical architectures, MAL allows the merging of heterogeneous architectures and selective parameters across layers. Specifically, the base model can incorporate parameters from different layers of multiple pre-trained models. We systematically investigate the conditions and fundamental settings of heterogeneous parameter merging, addressing all possible mismatches in layer widths between the base and target models. Furthermore, we establish key laws and provide practical guidelines for effectively implementing MAL.

In this study, we examine the potential of one of the ``superexpressive'' networks in the context of learning neural functions for representing complex signals and performing machine learning downstream tasks. Our focus is on evaluating their performance on computer vision and scientific machine learning tasks including signal representation/inverse problems and solutions of partial differential equations. Through an empirical investigation in various benchmark tasks, we demonstrate that superexpressive networks, as proposed by [Zhang et al. NeurIPS, 2022], which employ a specialized network structure characterized by having an additional dimension, namely width, depth, and ``height'', can surpass recent implicit neural representations that use highly-specialized nonlinear activation functions.

In this paper, we explore the idea of combining GCNs into one model. To that end, we align the weights of different models layer-wise using optimal transport (OT). We present and evaluate three types of transportation costs and show that the studied fusion method consistently outperforms the performance of vanilla averaging. Finally, we present results suggesting that model fusion using OT is harder in the case of GCNs than MLPs and that incorporating the graph structure into the process does not improve the performance of the method.

Foundation models for 3D shape generation have recently shown a remarkable capacity to encode rich geometric priors across both global and local dimensions. However, leveraging these priors for downstream tasks can be challenging as real-world data are often scarce or noisy, and traditional fine-tuning can lead to catastrophic forgetting. In this work, we treat the weight space of a large 3D shape-generative model as a data modality that can be explored directly. We hypothesize that submanifolds within this high-dimensional weight space can modulate topological properties or fine-grained part features separately, demonstrating early-stage evidence via two experiments. First, we observe a sharp phase transition in global connectivity when interpolating in conditioning space, suggesting that small changes in weight space can drastically alter topology. Second, we show that low-dimensional reparameterizations yield controlled local geometry changes even with very limited data. These results highlight the potential of weight space learning to unlock new approaches for 3D shape generation and specialized fine-tuning.

Mitigating Trojans in Large Language Models (LLMs) is one of many tasks where alignment data is LLM specific, as different LLMs have different Trojan triggers and trigger behaviors to be removed. In this paper, we introduce TeleLoRA (Teleporting Low-Rank Adaptation), a novel framework that synergizes model-specific alignment data across multiple LLMs to enable zero-shot Trojan mitigation on unseen LLMs without alignment data. TeleLoRA learns a unified generator of LoRA adapter weights by leveraging local activation information across multiple LLMs. This generator is designed to be permutation symmetric to generalize across models with different architectures and sizes. We optimize the model design for memory efficiency, making it feasible to learn with large-scale LLMs with minimal computational resources. Experiments on LLM Trojan mitigation benchmarks demonstrate that TeleLoRA effectively reduces attack success rates while preserving the benign performance of the models.

Neural quantum states (NQS) have emerged as a powerful tool for approximating quantum wavefunctions using deep learning. While these models achieve remarkable accuracy, understanding how they encode physical information remains an open challenge. In this work, we introduce adiabatic fine-tuning, a scheme that trains NQS across a phase diagram, leading to strongly correlated weight representations across different models. This correlation in weight space enables the detection of phase transitions in quantum systems by analyzing the trained network weights alone. Our results establish a connection between physical phase transitions and the geometry of neural network parameters, opening new directions for the interpretability of machine learning models in physics.