LoPT让大模型微调更省内存更快，只更新中间层后半部分。
Rethinking Local Learning: A Cheaper and Faster Recipe for LLM Post-Training
- 在Transformer中间设置梯度边界，只让后半层学任务目标。
- 前半层用轻量重建目标保持原始特征，减少对预训练表示的干扰。
- 相比传统方法，内存更低、速度更快，且更好保留预训练能力。
大模型微调通常将任务梯度传播至模型全部深度，导致需存储全层激活、存在长程反向依赖，并使任务梯度直接作用于预训练表征。我们指出这种全深度反向耦合可能过于昂贵且侵入性强，尤其当微调监督信号远少于预训练时。为此，我们提出**LoPT**：局部学习微调，将梯度路径限制在模型中点处。后半层直接学习任务目标，前半层则通过轻量特征重构目标维持有用表示并保持接口兼容性。该方法缩短了任务引发的反向传播路径，减少了窄任务梯度对浅层表示的直接干扰。大量实验表明，LoPT在保持竞争力性能的同时，显著降低内存开销、提升训练效率，并更好地保留预训练能力。代码已公开于：https://github.com/HumyuShi/LoPT。
原文摘要 · Abstract (English)
LLM post-training typically propagates task gradients through the full depth of the model. Although this end-to-end structure is simple and general, it couples task adaptation to full-depth activation storage, long-range backward dependencies and direct task-gradient access to pretrained representations. We argue that this full-depth backward coupling can be unnecessarily expensive and intrusive, particularly when post-training supervision is much narrower than pre-training. To this end, we propose \textbf{LoPT}: Local-Learning Post-Training, a simple post-training strategy that makes gradient reach an explicit design choice. LoPT places a single gradient boundary at the transformer midpoint: the second-half block learns from the task objective, while the first-half block is updated by a lightweight feature-reconstruction objective to preserve useful representations and maintain interface compatibility. LoPT shortens the task-induced backward path while limiting direct interference from narrow task gradients on early-layer representations. Extensive experiments demonstrate that LoPT achieves competitive performance with lower memory cost, higher training efficiency and better retention of pretrained capabilities. Our code is available at: https://github.com/HumyuShi/LoPT
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。