On October 24, 2026, a paper titled “Verbalizing Activation Differences with Natural Language Autoencoders” was published on the arXiv platform (arXiv:2610.30934v1), focusing on the ability of natural language autoencoders (NLAs) to provide human-readable representations of internal activation states in large language models. NLAs reconstruct model activations via a natural-language bottleneck, converting high-dimensional vectors into text descriptions that are understandable to humans.
The study conducts empirical analyses around four core questions: whether NLA-generated descriptions exhibit input compositionality; whether they can generalize to fine-tuned versions of the original model; whether they can recover the fine-tuning objective in model diffing; and whether they can be specifically optimized for diffing tasks. The experiments cover large language models with four distinct architectures, and the results broadly support positive conclusions: for instance, in analogy tasks, NLA descriptions predominantly point to the target word while rarely mentioning the subtracted term; moreover, the same NLA can be directly reused on lightweight fine-tuned models, with only a noticeable drop in reconstruction quality observed on the largest model.
In the model‑difference setting, NLA verbalizes the activation differences between “taboo model organisms” and their original models, successfully recovering the hidden key terms in most cases; the introduction of reinforcement learning further enhances its ability to recover key terms for unseen organisms. Notably, NLA can identify the positions of critical tokens preceding the prompt and reconstruct the fine-tuned target without relying on any output tokens generated by a fine-tuned model. Although it has yet to outperform the current state-of-the-art differential methods on the emerging task of emergent misalignment detection, it already significantly surpasses black-box baselines.