Optimal Sparsity of Mixture-of-Experts Language Models for Reasoning Tasks

• URL: http://arxiv.org/abs/2508.18672v1
• Date: Tue, 26 Aug 2025 04:31:28 GMT
• Title: Optimal Sparsity of Mixture-of-Experts Language Models for Reasoning Tasks
• Authors: Taishi Nakamura, Satoki Ishikawa, Masaki Kawamura, Takumi Okamoto, Daisuke Nohara, Jun Suzuki, Rio Yokota,
• Abstract summary: Mixture-of-Experts (MoE) models are now standard in state-of-the-art systems.<n>We investigate how MoE sparsity influences two distinct capability regimes: memorization and reasoning.
• Score: 17.067788440109137
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Empirical scaling laws have driven the evolution of large language models (LLMs), yet their coefficients shift whenever the model architecture or data pipeline changes. Mixture-of-Experts (MoE) models, now standard in state-of-the-art systems, introduce a new sparsity dimension that current dense-model frontiers overlook. We investigate how MoE sparsity influences two distinct capability regimes: memorization and reasoning. We train families of MoE Transformers that systematically vary total parameters, active parameters, and top-$k$ routing while holding the compute budget fixed. For every model we record pre-training loss, downstream task loss, and task accuracy, allowing us to separate the train-test generalization gap from the loss-accuracy gap. Memorization benchmarks improve monotonically with total parameters, mirroring training loss. By contrast, reasoning performance saturates and can even regress despite continued gains in both total parameters and training loss. Altering top-$k$ alone has little effect when active parameters are constant, and classic hyperparameters such as learning rate and initialization modulate the generalization gap in the same direction as sparsity. Neither post-training reinforcement learning (GRPO) nor extra test-time compute rescues the reasoning deficit of overly sparse models. Our model checkpoints, code and logs are open-source at https://github.com/rioyokotalab/optimal-sparsity.

Related papers

• Scaling Law Analysis in Federated Learning: How to Select the Optimal Model Size? [12.791994483385409]
  Concerns are growing about the depletion of high-quality, well-curated training data.<n>The decentralization of training datasets in Federated Learning introduces challenges to scaling large models.<n>This paper provides insights on generalizing the previous model scaling experience to federated learning scenarios.
  arXiv  Detail & Related papers  (2025-11-15T12:41:25Z)
• CAST: Continuous and Differentiable Semi-Structured Sparsity-Aware Training for Large Language Models [27.682531424487564]
  Sparsity-aware training is an effective approach for transforming large language models into hardware-friendly sparse patterns.<n>We propose Continuous Adaptive Sparse Trainer (CAST), a continuous and differentiable sparsity-aware training framework for sparse models.<n>Our results demonstrate significant improvements over previous state-of-the-art methods in both perplexity and zero-shot accuracy with minimal training resources.
  arXiv  Detail & Related papers  (2025-09-30T09:28:47Z)
• Compute-Optimal Scaling for Value-Based Deep RL [99.680827753493]
  We investigate compute scaling for online, value-based deep RL.<n>Our analysis reveals a nuanced interplay between model size, batch size, and UTD.<n>We provide a mental model for understanding this phenomenon and build guidelines for choosing batch size and UTD.
  arXiv  Detail & Related papers  (2025-08-20T17:54:21Z)
• It Takes a Good Model to Train a Good Model: Generalized Gaussian Priors for Optimized LLMs [15.263422862969803]
  We introduce BackSlash, a training-time compression algorithm for large language models.<n>We propose a unified, end-to-end framework for LLM optimization based on the GG model.<n>Our contributions are threefold:.<n>DeepShape, a post-training regularization method that reshapes weight distributions to match a GG profile,.<n>RF8, a compact and hardware-efficient 8-bit floating-point format designed for GG-distributed-priord BackSlash training.
  arXiv  Detail & Related papers  (2025-05-31T09:49:17Z)
• Don't be lazy: CompleteP enables compute-efficient deep transformers [50.85418589942566]
  Some parameterizations fail to transfer optimal base HPs across changes in model depth.<n>We develop theory to show parameterizations may still exist in the lazy learning regime.<n>We identify and adopt the parameterization we call CompleteP that achieves both depth-wise HP transfer and non-lazy learning in all layers.
  arXiv  Detail & Related papers  (2025-05-02T22:45:14Z)
• Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretraining [74.83412846804977]
  Reinforcement learning (RL)-based fine-tuning has become a crucial step in post-training language models.<n>We present a systematic end-to-end study of RL fine-tuning for mathematical reasoning by training models entirely from scratch.
  arXiv  Detail & Related papers  (2025-04-10T17:15:53Z)
• Do Larger Language Models Generalize Better? A Scaling Law for Implicit Reasoning at Pretraining Time [73.22651918134808]
  This work shows counterintuitive effects of model size scaling and provides new insights into the relationship between scaling and reasoning in language models (LMs)<n>We pretrain LMs from scratch on a synthetic implicit multihop reasoning environment designed to replicate the structure and distribution of real-world large-scale knowledge graphs.<n>We then assess the LMs' ability to complete the missing edges in the graph, which requires multi-hop reasoning that can be viewed as a simplification of implicit reasoning during real-world pretraining.
  arXiv  Detail & Related papers  (2025-04-04T17:57:22Z)
• LESA: Learnable LLM Layer Scaling-Up [57.0510934286449]
  Training Large Language Models (LLMs) from scratch requires immense computational resources, making it prohibitively expensive.<n>Model scaling-up offers a promising solution by leveraging the parameters of smaller models to create larger ones.<n>We propose textbfLESA, a novel learnable method for depth scaling-up.
  arXiv  Detail & Related papers  (2025-02-19T14:58:48Z)
• Parameters vs FLOPs: Scaling Laws for Optimal Sparsity for Mixture-of-Experts Language Models [10.517704202614091]
  sparse Mixture-of-Experts (MoEs) allow scaling the number of parameters without proportionally increasing the FLOPs per example.<n>We investigate how varying the sparsity level, i.e., the fraction of inactive parameters, impacts model's performance during pretraining and downstream few-shot evaluation.
  arXiv  Detail & Related papers  (2025-01-21T18:51:15Z)
• What Do Learning Dynamics Reveal About Generalization in LLM Reasoning? [83.83230167222852]
  We find that a model's generalization behavior can be effectively characterized by a training metric we call pre-memorization train accuracy. By connecting a model's learning behavior to its generalization, pre-memorization train accuracy can guide targeted improvements to training strategies.
  arXiv  Detail & Related papers  (2024-11-12T09:52:40Z)
• Scaling Laws for Sparsely-Connected Foundation Models [70.41266138010657]
  We explore the impact of parameter sparsity on the scaling behavior of Transformers trained on massive datasets. We identify the first scaling law describing the relationship between weight sparsity, number of non-zero parameters, and amount of training data.
  arXiv  Detail & Related papers  (2023-09-15T16:29:27Z)
• Sparse Model Soups: A Recipe for Improved Pruning via Model Averaging [24.64264715041198]
  We introduce Sparse Model Soups (SMS), a novel method for merging sparse models by initiating each prune-retrain cycle with the averaged model from the previous phase. SMS preserves sparsity, exploits sparse network benefits, is modular and fully parallelizable, and substantially improves IMP's performance.
  arXiv  Detail & Related papers  (2023-06-29T08:49:41Z)
• Winner-Take-All Column Row Sampling for Memory Efficient Adaptation of Language Model [89.8764435351222]
  We propose a new family of unbiased estimators called WTA-CRS, for matrix production with reduced variance. Our work provides both theoretical and experimental evidence that, in the context of tuning transformers, our proposed estimators exhibit lower variance compared to existing ones.
  arXiv  Detail & Related papers  (2023-05-24T15:52:08Z)
• Theoretical Characterization of the Generalization Performance of Overfitted Meta-Learning [70.52689048213398]
  This paper studies the performance of overfitted meta-learning under a linear regression model with Gaussian features. We find new and interesting properties that do not exist in single-task linear regression. Our analysis suggests that benign overfitting is more significant and easier to observe when the noise and the diversity/fluctuation of the ground truth of each training task are large.
  arXiv  Detail & Related papers  (2023-04-09T20:36:13Z)
• DSEE: Dually Sparsity-embedded Efficient Tuning of Pre-trained Language Models [152.29364079385635]
  As pre-trained models grow bigger, the fine-tuning process can be time-consuming and computationally expensive. We propose a framework for resource- and parameter-efficient fine-tuning by leveraging the sparsity prior in both weight updates and the final model weights. Our proposed framework, dubbed Dually Sparsity-Embedded Efficient Tuning (DSEE), aims to achieve two key objectives: (i) parameter efficient fine-tuning and (ii) resource-efficient inference.
  arXiv  Detail & Related papers  (2021-10-30T03:29:47Z)

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.