Meta 对比 Muse Wonder Full 与 Muse Spark 1.3 的测试时计算扩展表现
Meta 发了组对比实验：Muse Wonder Full 在没见过的任务上比 Muse Spark 1.3 预测的更好，他们把原因追溯到更强的基座模型，做模型选型可以看看。
Meta 用测试时计算扩展实验来衡量模型应对陌生任务的能力，对比了后训练见过的任务与未见任务。两款模型在分布外任务上的提升幅度相近，但在相同分布内分数下，Muse Wonder Full 在陌生任务上的表现优于 Muse Spark 1.3 的趋势预测。Meta 将这一上移归因于更强的基座模型，使其能把后训练收益更有效地迁移到未预料到的任务上。
To quantify how the models handle unfamiliar tasks, we compared test-time compute scaling on tasks trained on during post-training against other tasks. Both models improve similarly out-of-distribution, but given the same in-distribution score, Muse Wonder Full performs better on unfamiliar tasks than the Muse Spark 1.3 trend predicts. We attribute this upward shift to a stronger base model, which carries post-training gains over more effectively to tasks we didn't anticipate. 💬 2 🔄 0 ❤️ 1 👀 549 📊 2 ⚡
