J-space comparisons across open models
This article extends previous research on 'J-space'—how a model's middle layers steer its output—from closed models to open models. It investigates the temporal reach of steering, training formation, transferability between models, and scaling behavior, utilizing an autonomous agent for experimentation. The author shares the results in a raw, 'vibe coded' format to quickly disseminate findings.
Why it matters
Understanding J-space could lead to more interpretable, controllable, and efficient large language models, potentially enabling better fine-tuning, debugging, and safety mechanisms in AI. The methodology also highlights the increasing role of AI agents in accelerating scientific research.
Anthropic's Verbalizable-Workspace paper showed, on one closed model family, that a model's middle layers carry a dictionary of directions that causally steer its output. It left the natural next questions open: how far forward in time the steering reaches, when the structure forms during training, whether it transfers between models, and how it scales. We measured all four on open models, then two follow-ups the results forced on us. Every number below was re-derived from the committed result files, and every chart is interactive.
The article presents research findings and methodology in the field of AI, focusing on technical and scientific aspects without any political or ideological bias.
Get smarter about the news
Sign up free for a feed built around what you actually care about, Dive Deeper research on any story, and the full text of every article.
Create free accountAlready have an account? Sign in