arXiv subjects

Bedionita Soro

Publications and source records attributed to Bedionita Soro.

3 recordsLinked to original sources

arXiv2026

SLEM: Spectrum-Aware Latent Expert Merging

Mixture-of-Experts (MoE) models scale model capacity efficiently by activating only a small subset of experts for each input, while still requiring all parameters to be stored in memory, resulting in substantial memory overhead. Expert merging techniques have recently emerged as a promising compression paradigm that reduces this overhead by consolidating multiple experts into a smaller set. However, existing approaches have two key limitations: (i) direct convex interpolation in parameter space is not generally guaranteed to yield meaningful interpolation of expert functions; and (ii) estimating expert similarity or merging coefficients often relies on calibration data, introducing dependence on the calibration distribution. To address these limitations, we propose Spectrum-Aware Latent Expert Merging (SLEM), a calibration-free expert merging framework that occurs in a structured latent representation of expert weights using a VAE. SLEM groups experts within and across layers based on latent similarities and derives merging coefficients from expert weight spectra. It then performs geometry-aware latent interpolation and decodes the resulting representations into merged expert weights. Compared with existing expert merging methods, SLEM achieves comparable or better performance on general knowledge benchmarks and higher average performance on generation benchmarks involving mathematical reasoning and code generation.

cs.LG↗

arXiv2026

Distribution-Consistent Inference for Dynamic Sparse Mixture-of-Experts

Mixture-of-Experts (MoE) architectures have emerged as a powerful paradigm for scaling model capacity while preserving efficient inference in large foundation models. However, most MoE models use a fixed top-$k$ expert selection policy, assigning the same expert budget to every token even when fewer experts may be sufficient. Inference-time dynamic top-$k$ routing can reduce computation without retraining, but existing methods often overlook the distributional shift caused by deviating from the training-time routing configuration. We show that reducing the number of activated experts consistently increases the RMS scale and variance of SMoE outputs, inducing a representation mismatch that contributes to downstream performance degradation in addition to the loss of expert capacity. To address this correctable component, we propose Layer-wise Distribution Alignment (LDA), a lightweight inference-time correction that uses layer-wise calibration statistics to align reduced-routing representations with the default configuration. Across multiple SMoE LLMs, benchmarks, and routing strategies, LDA recovers much of the performance lost induced by the distributional shift under reduced routing while preserving sparse-inference efficiency with negligible overhead.

cs.LG↗

arXiv2024

Diffusion-Based Neural Network Weights Generation

Transfer learning has gained significant attention in recent deep learning research due to its ability to accelerate convergence and enhance performance on new tasks. However, its success is often contingent on the similarity between source and target data, and training on numerous datasets can be costly, leading to blind selection of pretrained models with limited insight into their effectiveness. To address these challenges, we introduce D2NWG, a diffusion-based neural network weights generation technique that efficiently produces high-performing weights for transfer learning, conditioned on the target dataset. Our method extends generative hyper-representation learning to recast the latent diffusion paradigm for neural network weights generation, learning the weight distributions of models pretrained on various datasets. This allows for automatic generation of weights that generalize well across both seen and unseen tasks, outperforming state-of-the-art meta-learning methods and pretrained models. Moreover, our approach is scalable to large architectures such as large language models (LLMs), overcoming the limitations of current parameter generation techniques that rely on task-specific model collections or access to original training data. By modeling the parameter distribution of LLMs, D2NWG enables task-specific parameter generation without requiring additional fine-tuning or large collections of model variants. Extensive experiments show that our method consistently enhances the performance of diverse base models, regardless of their size or complexity, positioning it as a robust solution for scalable transfer learning.

cs.LG↗