Optimizer-Model Consistency: Full Finetuning with the Same Optimizer as Pretraining Forgets Less

• URL: http://arxiv.org/abs/2605.06654v1
• Date: Thu, 07 May 2026 17:57:02 GMT
• Title: Optimizer-Model Consistency: Full Finetuning with the Same Optimizer as Pretraining Forgets Less
• Abstract summary: Full finetuning with the same as in pretraining achieves a better learning-forgetting tradeoff, i.e. forgetting less while achieving the same or better performance on the new task.<n>We use a synthetic language modeling experiment to show that this can come from Muon's strong tendency towards rote.
• Score: 20.230676271699743
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Optimizers play an important role in both pretraining and finetuning stages when training large language models (LLMs). In this paper, we present an observation that full finetuning with the same optimizer as in pretraining achieves a better learning-forgetting tradeoff, i.e., forgetting less while achieving the same or better performance on the new task, than other optimizers and, possibly surprisingly, LoRA, during the supervised finetuning (SFT) stage. We term this phenomenon optimizer-model consistency. To better understand it, through controlled experiments and theoretical analysis, we show that: 1) optimizers can shape the models by having regularization effects on the activations, leading to different landscapes around the pretrained checkpoints; 2) in response to this regularization effect, the weight update in SFT should follow some specific structures to lower forgetting of the knowledge learned in pretraining, which can be obtained by using the same optimizer. Moreover, we specifically compare Muon and AdamW when they are employed throughout the pretraining and SFT stages and find that Muon performs worse when finetuned for reasoning tasks. With a synthetic language modeling experiment, we demonstrate that this can come from Muon's strong tendency towards rote memorization, which may hurt pattern acquisition with a small amount of data, as for SFT.

Related papers

• A Defense of the Quadratic Model [71.57026347416729]
  We stress test the simplest possible model of optimization -- the quadratic model.<n>We show that it can be predictive surprisingly in an LLM setting with 150M parameters and 3B training tokens.
  arXiv  Detail & Related papers  (2026-07-23T18:02:03Z)
• (How) Learning Rates Regulate Catastrophic Overtraining [31.988459260329353]
  Supervised fine-tuning (SFT) is a common first stage of LLM post-training.<n>We investigate catastrophic forgetting in finetuning through the lens of implicit regularization of the learning rate.<n>We link forgetting to overtraining: learning rate decay increases the sharpness of the pretrained model, which in turn exacerbates catastrophic forgetting during SFT.
  arXiv  Detail & Related papers  (2026-04-15T08:53:42Z)
• REG: A Regularization Optimizer for Robust Training Dynamics [24.850151895583494]
  Row-and-Column-Scaling (RACS) operator regularizes the update steps in a less drastic manner, making it simpler to implement and more compatible with established training dynamics.<n>We demonstrate that our REG achieves superior performance and stability over AdamW, but also maintains consistency with the AdamW training paradigm.
  arXiv  Detail & Related papers  (2025-10-04T06:05:57Z)
• Overtrained Language Models Are Harder to Fine-Tune [64.44743256512237]
  Large language models are pre-trained on ever-growing token budgets.<n>We show that extended pre-training can make models harder to fine-tune, leading to degraded final performance.
  arXiv  Detail & Related papers  (2025-03-24T23:11:56Z)
• Muon is Scalable for LLM Training [50.68746986439438]
  We introduce Moonlight, a Mixture-of-Expert (MoE) model trained with 5.7T tokens using Muon.<n>Our model improves the current frontier, achieving better performance with much fewer training FLOPs compared to prior models.<n>We open-source our distributed Muon implementation that is memory optimal and communication efficient.
  arXiv  Detail & Related papers  (2025-02-24T09:12:29Z)
• PROFIT: A Specialized Optimizer for Deep Fine Tuning [8.303565426823674]
  We present PROFIT, one of the first designed to incrementally fine-tune converged models on new tasks and/or datasets.<n> PROFIT takes the properties of a converged model into account explicitly to regularize the optimization process.<n>It outperforms fine-tuning methods in various tasks, from image classification to multimodal language model training to large-scale motion prediction.
  arXiv  Detail & Related papers  (2024-12-02T19:37:34Z)
• Unlearning as multi-task optimization: A normalized gradient difference approach with an adaptive learning rate [105.86576388991713]
  We introduce a normalized gradient difference (NGDiff) algorithm, enabling us to have better control over the trade-off between the objectives.<n>We provide a theoretical analysis and empirically demonstrate the superior performance of NGDiff among state-of-the-art unlearning methods on the TOFU and MUSE datasets.
  arXiv  Detail & Related papers  (2024-10-29T14:41:44Z)
• HFT: Half Fine-Tuning for Large Language Models [42.60438623804577]
  Large language models (LLMs) with one or more fine-tuning phases have become a necessary step to unlock various capabilities. In this paper, we find that by regularly resetting partial parameters, LLMs can restore some of the original knowledge. We introduce Half Fine-Tuning (HFT) for LLMs, as a substitute for full fine-tuning (FFT), to mitigate the forgetting issues.
  arXiv  Detail & Related papers  (2024-04-29T07:07:58Z)
• 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)
• FTFT: Efficient and Robust Fine-Tuning by Transferring Training Dynamics [7.58472343957521]
  We show that training dynamics are highly transferable across model sizes and pre-training methods.<n>We propose a novel fine-tuning approach: Fine-Tuning by transFerring Training dynamics (FTFT)
  arXiv  Detail & Related papers  (2023-10-10T12:53:48Z)
• Learning to Optimize for Reinforcement Learning [58.01132862590378]
  Reinforcement learning (RL) is essentially different from supervised learning, and in practice, these learneds do not work well even in simple RL tasks. Agent-gradient distribution is non-independent and identically distributed, leading to inefficient meta-training. We show that, although only trained in toy tasks, our learned can generalize unseen complex tasks in Brax.
  arXiv  Detail & Related papers  (2023-02-03T00:11:02Z)

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.