Interpretability in AI has long been viewed as a trade-off: models are trained as black boxes, then explained post-hoc with methods of questionable reliability. A new research paper from Hugging Face challenges this assumption by integrating interpretability directly into the training pipeline, optimizing it alongside the language modeling objective.
The study, titled "Scaling Inherently Interpretable Language Models," demonstrates that 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 as scale increases.
The authors instantiate their 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 a closed-loop intervention: diagnose an output through concept or feature attribution, retrieve similar training data, and correct behavior through concept steering—all without retraining.
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.