Scaling Inherently Interpretable Language Models

扩展内在可解释语言模型

Interpretability is often treated as a tax on capability: language models are trained as opaque systems, then explained after the fact, with methods whose reliability is difficult to establish. In this work, we challenge this premise. Rather than reverse-engineering a model, we make interpretability a constraint of the training pipeline, optimized alongside the language modeling objective.

可解释性通常被视为性能的“税收”：语言模型被训练为不透明的系统，随后再通过难以验证其可靠性的方法进行事后解释。在这项工作中，我们挑战了这一前提。我们不再对模型进行逆向工程，而是将可解释性作为训练流程的一个约束条件，与语言建模目标同步优化。

Across three orders of magnitude of compute, on both autoregressive and diffusion language models, interpretability scales with capability rather than against it. Surprisingly, model representations become more disentangled and aligned with human-understandable concepts with scale.

在三个数量级的计算规模下，无论是自回归语言模型还是扩散语言模型，可解释性都随着性能的提升而增强，而非受其阻碍。令人惊讶的是，随着规模的扩大，模型表示变得更加解耦，并与人类可理解的概念更加一致。

We instantiate the training-time recipe with Steerling-8B, a diffusion language model with a causal attention mask. For any group of generated tokens, Steerling-8B attributes the output to relevant input tokens, human-understandable concepts, and training data. This enables closed-loop intervention: diagnose an output through its concept or feature attribution, retrieve similar training data, and correct the behavior through concept steering without retraining.

我们通过 Steerling-8B（一种带有因果注意力掩码的扩散语言模型）实现了这一训练方案。对于生成的任何一组 Token，Steerling-8B 都能将输出归因于相关的输入 Token、人类可理解的概念以及训练数据。这实现了闭环干预：通过概念或特征归因诊断输出，检索相似的训练数据，并在无需重新训练的情况下通过概念引导（Concept Steering）纠正模型行为。

Steerling-8B remains competitive with open peer models trained on substantially 2-16x more compute, suggesting a different scaling paradigm: interpretability can be designed into training, and it improves with scale.

Steerling-8B 的表现与那些计算资源投入多出 2 到 16 倍的开源同类模型相比依然具有竞争力，这表明了一种不同的扩展范式：可解释性可以被设计进训练过程中，并随着规模的扩大而提升。