Dense2MoE: Pushing the Pareto Frontier of On-Device LLMs via Unified Pruning and Upcycling

• URL: http://arxiv.org/abs/2605.26496v1
• Date: Tue, 26 May 2026 03:19:04 GMT
• Title: Dense2MoE: Pushing the Pareto Frontier of On-Device LLMs via Unified Pruning and Upcycling
• Abstract summary: Mixture of Experts MoE architecture is highly promising for resource constrained on device deployments yet training these models from scratch incurs prohibitive costs.<n>Current methods attempt to alleviate this by upcycling dense models into MoEs however they often introduce parameter redundancy that degrades inference efficiency Alternatively standard layer prunings redundancy but inevitably compromises model accuracy.
• Score: 10.009897539610646
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: The Mixture of Experts MoE architecture is highly promising for resource constrained on device deployments yet training these models from scratch incurs prohibitive costs Current methods attempt to alleviate this by upcycling dense models into MoEs however they often introduce parameter redundancy that degrades inference efficiency Alternatively standard layer pruning mitigates redundancy but inevitably compromises model accuracy To resolve this dilemma we propose Dense2MoE a novel framework that unifies pruning and upcycling through Layer Fusion UpCycling LF UC Guided by hardware Roofline theory Dense2MoE systematically overcomes the inference memory wall by pruning bandwidth heavy attention modules from redundant layers while repurposing their Multi Layer Perceptrons MLPs into MoE experts This structural innovation preserves the models core capabilities and strictly limits active parameters via selective token routing With a modest continual pre training budget Dense2MoE efficiently converts publicly available dense LLMs into on device ready MoE models Extensive experiments demonstrate that Dense2MoE significantly advances the Pareto frontier for on device inference latency versus model accuracy outperforming dense baselines state of the art compression and standard upcycling methods

Related papers

• When Model Merging Breaks Routing: Training-Free Calibration for MoE [67.27839627141972]
  Hessian-Aware Router (HARC) is a training-free framework that leverages second-order curvature information to realign the merged router.<n> Experiments on mathematical reasoning and code generation tasks show that HARC effectively mitigates routing breakdown across diverse MoE merging baselines.
  arXiv  Detail & Related papers  (2026-06-02T09:33:33Z)
• Self-Distilled Trajectory-Aware Boltzmann Modeling: Bridging the Training-Inference Discrepancy in Diffusion Language Models [65.89572755202245]
  Diffusion Language Models (DLMs) offer stronger global awareness and highly parallel generation.<n>Post-training DLMs with standard Negative Evidence Lower Bound (NELBO)-based supervised fine-tuning remains inefficient.<n>We propose a self-distilled trajectory-based post-training framework that aligns training with the easy-to-hard structure of inference.
  arXiv  Detail & Related papers  (2026-05-12T09:39:06Z)
• Efficient Training of Diffusion Mixture-of-Experts Models: A Practical Recipe [51.26601054313749]
  Recent efforts on Diffusion MoE models have primarily focused on developing more sophisticated routing mechanisms.<n>Inspired by the MoE design paradigms established in large language models (LLMs), we identify a set of crucial architectural factors for building effective Diffusion MoE models.<n>We present novel architectures that can be efficiently applied to both latent and pixel-space diffusion frameworks.
  arXiv  Detail & Related papers  (2025-12-01T03:52:31Z)
• MoE-Prism: Disentangling Monolithic Experts for Elastic MoE Services via Model-System Co-Designs [17.827406818899536]
  MoE-Prism is a model-system co-design that transforms rigid MoE models into elastic services.<n>Our evaluation shows that MoE-Prismprovides over 4 times more distinct, stable operating points than the baseline.<n>This allows an AI service to dynamically improve throughput by up to 19.9% under a strict budget or reduce latency by up to 10.36% under limited resources.
  arXiv  Detail & Related papers  (2025-10-22T08:40:01Z)
• Minitron-SSM: Efficient Hybrid Language Model Compression through Group-Aware SSM Pruning [76.88243649182886]
  Hybrid architectures that combine Attention and State Space Models (SSMs) achieve state-of-the-art accuracy and runtime performance.<n>Recent work has demonstrated that applying compression and distillation to Attention-only models yields smaller, more accurate models at a fraction of the training cost.<n>We introduce a novel group-aware pruning strategy that preserves the structural integrity of SSM blocks and their sequence modeling capabilities.
  arXiv  Detail & Related papers  (2025-04-15T17:26:29Z)
• S'MoRE: Structural Mixture of Residual Experts for Parameter-Efficient LLM Fine-tuning [19.038272193170297]
  We propose Structural Mixture of Residual Experts (S'MoRE), a novel framework that seamlessly integrates the efficiency of LoRA with the flexibility of MoE.<n>By routing input tokens through sub-trees of residuals, S'MoRE emulates the capacity of numerous experts by instantiating and assembling just a few low-rank matrices.
  arXiv  Detail & Related papers  (2025-04-08T20:54:00Z)
• Speculative MoE: Communication Efficient Parallel MoE Inference with Speculative Token and Expert Pre-scheduling [3.529891364583952]
  MoE (Mixture of Experts) prevails as a neural architecture that can scale modern transformer-based LLMs (Large Language Models) to unprecedented scales.<n>DeepSpeed-MoE, one state-of-the-art MoE inference framework, adopts a 3D-parallel paradigm including EP (Expert Parallelism), TP (Tensor Parallel) and DP (Data Parallelism)<n>Our work aims to boost DeepSpeed-MoE by strategically reducing EP's communication overhead with a technique named Speculative MoE.
  arXiv  Detail & Related papers  (2025-03-06T12:52:22Z)
• ToMoE: Converting Dense Large Language Models to Mixture-of-Experts through Dynamic Structural Pruning [24.8038863056542]
  Large Language Models (LLMs) have demonstrated remarkable abilities in tackling a wide range of complex tasks.<n>Their huge computational and memory costs raise significant challenges in deploying these models on resource-constrained devices.<n>We introduce a different dynamic pruning method that pushes dense models to maintain a fixed number of active parameters.
  arXiv  Detail & Related papers  (2025-01-25T20:01:42Z)
• HOBBIT: A Mixed Precision Expert Offloading System for Fast MoE Inference [54.40808356999408]
  We present HOBBIT, a mixed precision expert offloading system to enable flexible and efficient MoE inference. Our key insight is that dynamically replacing less critical cache-miss experts with low precision versions can substantially reduce expert-loading latency. HOBBIT achieves up to a 9.93x speedup in decoding compared to state-of-the-art MoE offloading systems.
  arXiv  Detail & Related papers  (2024-11-03T04:25:46Z)
• Read-ME: Refactorizing LLMs as Router-Decoupled Mixture of Experts with System Co-Design [59.00758127310582]
  We propose a novel framework Read-ME that transforms pre-trained dense LLMs into smaller MoE models. Our approach employs activation sparsity to extract experts. Read-ME outperforms other popular open-source dense models of similar scales.
  arXiv  Detail & Related papers  (2024-10-24T19:48:51Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.