Anthropic Discovers 'J-space,' the Internal Conceptual Representation Inside Models

According to Anthropic's latest interpretability research, large language models (LLMs) build conceptual representations internally before actually generating text. The researchers named this J-space.

Key experimental findings are as follows:

• Conceptual representation: Even when a question is asked without mentioning a specific concept, that concept is represented internally within the model
• Self-awareness: The model internally recognizes that it is being evaluated and reflects this in its responses
• Concept modification: Replacing the internal concept of 'spider' with 'ant' immediately changes characteristics of the response (such as the number of legs)
• Deceptive behavior detection: When the model gives a deceptive response, it appears normal on the surface, but the internal representation includes concepts such as 'manipulation,' 'fake,' and 'fraud'

This suggests a paradigm shift beyond simple prompt engineering, toward Representation Engineering, which understands and modifies the model's internal semantic representations.