Less is More: Undertraining Experts Improves Model Upcycling

• URL: http://arxiv.org/abs/2506.14126v1
• Date: Tue, 17 Jun 2025 02:42:10 GMT
• Title: Less is More: Undertraining Experts Improves Model Upcycling
• Authors: Stefan Horoi, Guy Wolf, Eugene Belilovsky, Gintare Karolina Dziugaite,
• Abstract summary: We show how expert fine-tuning affects model upcycling.<n>We trace this degradation to the memorization of a small set of difficult examples.<n>We demonstrate that a task-dependent aggressive early stopping strategy can significantly improve upcycling performance.
• Score: 28.79396346914521
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Modern deep learning is increasingly characterized by the use of open-weight foundation models that can be fine-tuned on specialized datasets. This has led to a proliferation of expert models and adapters, often shared via platforms like HuggingFace and AdapterHub. To leverage these resources, numerous model upcycling methods have emerged, enabling the reuse of fine-tuned models in multi-task systems. A natural pipeline has thus formed to harness the benefits of transfer learning and amortize sunk training costs: models are pre-trained on general data, fine-tuned on specific tasks, and then upcycled into more general-purpose systems. A prevailing assumption is that improvements at one stage of this pipeline propagate downstream, leading to gains at subsequent steps. In this work, we challenge that assumption by examining how expert fine-tuning affects model upcycling. We show that long fine-tuning of experts that optimizes for their individual performance leads to degraded merging performance, both for fully fine-tuned and LoRA-adapted models, and to worse downstream results when LoRA adapters are upcycled into MoE layers. We trace this degradation to the memorization of a small set of difficult examples that dominate late fine-tuning steps and are subsequently forgotten during merging. Finally, we demonstrate that a task-dependent aggressive early stopping strategy can significantly improve upcycling performance.

Related papers

• GRAM: A Generative Foundation Reward Model for Reward Generalization [48.63394690265176]
  We develop a generative reward model that is first trained via large-scale unsupervised learning and then fine-tuned via supervised learning.<n>This model generalizes well across several tasks, including response ranking, reinforcement learning from human feedback, and task adaptation with fine-tuning.
  arXiv  Detail & Related papers  (2025-06-17T04:34:27Z)
• Drop-Upcycling: Training Sparse Mixture of Experts with Partial Re-initialization [18.271311365080802]
  Mixture of Experts (MoE) architecture reduces the training and inference cost significantly compared to a dense model of equivalent capacity.<n>Upcycling is an approach that initializes and trains an MoE model using a pre-trained dense model.<n>Drop-Upcycling combines two seemingly contradictory approaches: utilizing the knowledge of pre-trained dense models while statistically re-initializing some parts of the weights.
  arXiv  Detail & Related papers  (2025-02-26T16:06:36Z)
• Scaling Laws for Upcycling Mixture-of-Experts Language Models [17.796361238003403]
  Pretraining large language models (LLMs) is resource-intensive, often requiring months of training time even with high-end GPU clusters.<n>There are two approaches of mitigating such computational demands: reusing smaller models to train larger ones (upcycling) and training computationally efficient models like mixture-of-experts (MoE)
  arXiv  Detail & Related papers  (2025-02-05T09:11:13Z)
• Meta-Learning Adaptable Foundation Models [37.458141335750696]
  We introduce a meta-learning framework infused with PEFT in this intermediate retraining stage to learn a model that can be easily adapted to unseen tasks. In this setting, we demonstrate the suboptimality of standard retraining for finding an adaptable set of parameters. We then apply these theoretical insights to retraining the RoBERTa model to predict the continuation of conversations within the ConvAI2 dataset.
  arXiv  Detail & Related papers  (2024-10-29T17:24:18Z)
• Preparing Lessons for Progressive Training on Language Models [75.88952808979087]
  The rapid progress of Transformers in artificial intelligence has come at the cost of increased resource consumption and greenhouse gas emissions. We propose Apollo, which preptextbfares lessons for extextbfpanding textbfoperations by textbflayer functitextbfonality during training of low layers. Experiments demonstrate that Apollo achieves state-of-the-art acceleration ratios, even rivaling methods using pretrained models.
  arXiv  Detail & Related papers  (2024-01-17T13:04:14Z)
• An Emulator for Fine-Tuning Large Language Models using Small Language Models [91.02498576056057]
  We introduce emulated fine-tuning (EFT), a principled and practical method for sampling from a distribution that approximates the result of pre-training and fine-tuning at different scales. We show that EFT enables test-time adjustment of competing behavioral traits like helpfulness and harmlessness without additional training. Finally, a special case of emulated fine-tuning, which we call LM up-scaling, avoids resource-intensive fine-tuning of large pre-trained models by ensembling them with small fine-tuned models.
  arXiv  Detail & Related papers  (2023-10-19T17:57:16Z)
• Continual Learners are Incremental Model Generalizers [70.34479702177988]
  This paper extensively studies the impact of Continual Learning (CL) models as pre-trainers. We find that the transfer quality of the representation often increases gradually without noticeable degradation in fine-tuning performance. We propose a new fine-tuning scheme, GLobal Attention Discretization (GLAD), that preserves rich task-generic representation during solving downstream tasks.
  arXiv  Detail & Related papers  (2023-06-21T05:26:28Z)
• Retrieval-Enhanced Contrastive Vision-Text Models [61.783728119255365]
  We propose to equip vision-text models with the ability to refine their embedding with cross-modal retrieved information from a memory at inference time. Remarkably, we show that this can be done with a light-weight, single-layer, fusion transformer on top of a frozen CLIP. Our experiments validate that our retrieval-enhanced contrastive (RECO) training improves CLIP performance substantially on several challenging fine-grained tasks.
  arXiv  Detail & Related papers  (2023-06-12T15:52:02Z)
• An Empirical Investigation of the Role of Pre-training in Lifelong Learning [21.995593026269578]
  We show that generic pre-training implicitly alleviates the effects of catastrophic forgetting when learning multiple tasks sequentially. We study this phenomenon by analyzing the loss landscape, finding that pre-trained weights appear to ease forgetting by leading to wider minima.
  arXiv  Detail & Related papers  (2021-12-16T19:00:55Z)
• Powerpropagation: A sparsity inducing weight reparameterisation [65.85142037667065]
  We introduce Powerpropagation, a new weight- parameterisation for neural networks that leads to inherently sparse models. Models trained in this manner exhibit similar performance, but have a distribution with markedly higher density at zero, allowing more parameters to be pruned safely. Here, we combine Powerpropagation with a traditional weight-pruning technique as well as recent state-of-the-art sparse-to-sparse algorithms, showing superior performance on the ImageNet benchmark.
  arXiv  Detail & Related papers  (2021-10-01T10:03:57Z)
• Extrapolation for Large-batch Training in Deep Learning [72.61259487233214]
  We show that a host of variations can be covered in a unified framework that we propose. We prove the convergence of this novel scheme and rigorously evaluate its empirical performance on ResNet, LSTM, and Transformer.
  arXiv  Detail & Related papers  (2020-06-10T08:22:41Z)

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.