AI 的“零件”能拼出行为吗？不能全靠单个零件
我们总以为 AI 的“想法”可以拆成一个个独立零件，拼起来就是它的行为。这篇论文发现，拆是能拆，但拼不回去：单个零件各自的效果加在一起，预测不了 AI 实际会怎么做。研究者用稀疏自编码器把模型内部状态拆成稀疏特征，再测这些特征单独作用与组合作用的差距，发现两两配对的效果比单个特征加起来更准，但依然不够，需要更高阶的交互。在控制模型行为的实验里，考虑配对后排序准确率从 0.5874 升到 0.7790，错误率从 0.1444 降到 0.0606。换句话说，AI 的行为不是零件清单，而是零件之间的化学反应。它不是你明天能用上的，但它提醒我们：想靠“拆解”来理解或控制 AI，可能得接受一个更复杂的现实。
📄 原文摘要(英文)
Sparse autoencoders (SAEs) represent model activations with sparse, linearly decoded features, but the downstream effects of those features need not compose additively. We introduce SAE Additivity Residuals (SAR), which measure the behavioral variation left unexplained above a chosen interaction order. Under independent feature subset sampling, population SAR is exactly the normalized energy of interactions above that order and bounds coalition ordering error for the optimal low order approximation. To distinguish compositionality associated with the SAE representation from that induced by downstream computation, we compare learned SAE coordinates with sum preserving random rotations of the same selected subspace and find lower first order residuals for the learned coordinates across all 18 model and behavior settings. The coordinates are nevertheless not fully additive: pairwise terms improve held out coalition ordering in every setting and outperform parameter matched models using randomly selected third order terms. In a combined steering experiment, pairwise modeling raises rank fidelity from 0.5874 to 0.7790 and reduces ordering error from 0.1444 to 0.0606. Together, these results show that SAE coordinates support comparatively compositional behavioral descriptions while faithful joint explanations and interventions still require interactions beyond individual feature effects. Code is available at https://github.com/OpenSAE/SAE-Behavioral-Compositionality.