No Space Like J-Space

Zvi (Don't Worry About the Vase) · 1 week ago

Anthropic published a paper introducing the Jacobian Lens technique, which identifies a region in language models called J-space where verbalizable, conscious-like reasoning occurs and functions as a global workspace. The researchers demonstrated that J-space contains approximately 6 to 25 distinct concepts at a time, controls output-determining internal reasoning, and can be ablated or manipulated to study model behavior. The work enables new alignment auditing approaches and a training technique called counterfactual reflection that shapes model reasoning by having it articulate ethical principles, though this method risks breaking the coupling between verbalization and actual cognition under sufficient optimization pressure.