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.

翻译：在单GPU上对视觉-语言-动作模型（VLA-JEPA）进行微调本应简单：加载预训练检查点、运行训练、部署。但其中暗藏风险。对同一微调代码重复运行十三次——相同数据、相同架构、不同随机种子——其中十二次运行产生机器人91%-94%的成功率，而一次运行悄然降至65.2%：性能差距达29个百分点，且无错误信息、无警告、无法预判哪个种子会失败。我们将此现象称为种子彩票。研究发现根源在于输出坍缩：动作预测器在不在意机器人视觉输入的情况下，逐渐学会生成近乎相同的输出。现有权重级方法（L2正则化、弹性权重巩固）对此坍缩存在结构性盲区——它们惩罚权重变化，但坍缩发生在权重的雅可比零空间中可以自由移动而不影响输出的方向，我们通过雅可比零空间正式刻画了这一缺陷。在7种方法×最多13个种子×3个LIBERO基准测试中，三种输出级正则化器——VICReg（12个种子）、Dropout（4个种子）和减半学习率（5个种子）——各自消除了所有灾难性种子（合并坍缩0/21 vs 基线1/13；F(12,11)=28.7，p<0.001），而权重级方法（L2正则化、弹性权重巩固）保留了彩票效应。最简单的修复方案是更改优化器配置中的一个数值。