arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2512.15134cs.LGcs.AIcs.CL
收藏

从孤立到纠缠:可解释性方法何时识别和解缠已知概念?

From Isolation to Entanglement: When Do Interpretability Methods Identify and Disentangle Known Concepts?

Boston University(波士顿大学);Harvard University(哈佛大学);Mila – Quebec AI Institute(魁北克AI研究所);Goodfire(Goodfire公司)

Aaron Mueller, Andrew Lee, Shruti Joshi, Ekdeep Singh Lubana, Dhanya Sridhar, Patrik Reizinger

AI总结:

本文提出多概念评估框架,研究稀疏自编码器和探针等方法是否真正解缠概念,发现特征通常只对单一概念敏感,但概念分布在多个特征上,且干预特征常影响多个概念,表明相关性指标不足以证明干预选择性。

摘要:

可解释性的一个目标是从神经网络的激活中恢复潜在概念(特征)的解缠表示。特征的质量通常孤立地评估,并在可能不成立的隐式独立性假设下进行。因此,尚不清楚常见的特征化方法(如稀疏自编码器(SAE)和探针)在多大程度上将一个概念与另一个概念解缠。我们提出了一个多概念评估设置,使用包括情感、领域、语态和时态在内的概念。我们评估特征化器产生每个概念的解缠表示的效果,观察到特征通常只对单一概念敏感,但概念分布在许多特征上。然后,我们干预这些特征,测量每个概念是否可独立操控,以及特征是否相互作用。即使在理想化设置中,干预一个特征通常会影响多个概念,尽管几乎没有交互效应。这些结果表明,相关性指标不足以建立干预选择性,并且证明两个特征在分离空间中运行不足以声称它们将对一个概念具有选择性。这些结果强调了可解释性研究中多概念评估的重要性。

英文摘要:

A goal of interpretability is to recover disentangled representations of latent concepts (features) from the activations of neural networks. The quality of features is typically evaluated in isolation, and under implicit independence assumptions that may not hold in practice. Thus, it is unclear to what extent common featurization methods such as sparse autoencoders (SAEs) and probes disentangle one concept from another. We propose a multi-concept evaluation setting using concepts including sentiment, domain, voice, and tense. We evaluate how well featurizers produce disentangled representations of each concept, observing that features are typically sensitive to only one concept, but also that concepts are distributed across many features. Then, we steer these features, measuring whether each concept is independently manipulable, and whether features interact. Even in idealized settings, steering a feature often affects many concepts, despite a near absence of interaction effects. These results suggest that correlational metrics are insufficient to establish steering selectivity, and that demonstrating that two features operate in separate spaces is insufficient to claim that they will be selective for one concept. These results underscore the importance of multi-concept evaluations in interpretability research.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑