单卡微调时，换随机种子可能让机器人成功率从94%暴跌到65%，新方法通过输出层正则化彻底解决此问题。
Output-Level Regularization Eliminates the Seed Lottery in Single-GPU VLA Fine-Tuning
- 在输出层引入正则化，防止动作预测器输出趋同
- 13次实验中仅1次失败（0/21崩溃），对比基线1/13失败率显著提升
- 只需调整优化器学习率等参数，适合实际部署场景
在单卡上微调视觉-语言-动作模型（VLA-JEPA）看似简单：加载预训练权重，运行训练，直接部署。但存在隐藏风险：相同数据与架构下，13次不同随机种子的实验中，12次成功率达91–94%，而1次无声退化至65.2%，下降29个百分点，无警告也无预警，无法预测失败种子。我们称此为‘种子彩票’。根源是输出坍缩：动作预测器逐渐输出几乎相同的动作，与输入无关。现有权重级正则化（如L2、EWC）对此类坍缩结构上失效——它们惩罚权重变化，但坍缩发生在权重可自由移动却不影响输出的方向上，该现象我们通过雅可比零空间形式化分析。在7种方法 × 最多13个种子 × 3个LIBERO基准测试中，三种输出级正则化（VICReg，n=12；Dropout，n=4；学习率减半，n=5）均消除所有灾难性失败（0/21崩溃，基线1/13；F(12,11)=28.7，p<0.001），而权重级方法仍保留彩票效应。最简单的修复只需更改优化器配置中的一个数值。
原文摘要 · Abstract (English)
Fine-tuning a vision-language-action model (VLA-JEPA) on a single GPU should be simple: load a pretrained checkpoint, run training, deploy. There is a hidden danger. Run the same fine-tuning code thirteen times -- same data, same architecture, different random seed -- and twelve runs produce a robot succeeding 91--94% of the time, while one run silently degrades to 65.2%: a 29 pp gap with no error message, no warning, and no way to predict which seed will fail. We call this the seed lottery. We trace the cause to output collapse: the action predictor quietly learns to produce nearly identical outputs regardless of what the robot sees. Existing weight-level methods (L2, EWC) are structurally blind to this collapse -- they penalize weight changes, but collapse occurs in directions weights can move freely without affecting outputs, a gap we formalize via the Jacobian null-space. Across 7 methods x up to 13 seeds x 3 LIBERO benchmarks, three output-level regularizers -- VICReg (n=12 seeds), Dropout (n=4), and a halved learning rate (n=5) -- each eliminate every catastrophic seed (0/21 combined collapses vs. 1/13 Baseline; F(12,11)=28.7, p<0.001), while weight-level methods (L2, EWC) preserve the lottery. The simplest fix is changing one number in your optimizer config.
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。