[From 'Dictionary Features' to 'Stream of Thought Monitoring': Anthropic's Jacobian Space Research Unveils the Brain-like Reasoning Hub in Large Models] According to Beating's monitoring, research on the mechanistic interpretability of large models is evolving from 'static dictionaries' to 'dynamic streams of thought.' Previously, Anthropic used sparse autoencoders (SAE) to disentangle millions of isolated conceptual features (e.g., 'Golden Gate Bridge') within Claude, effectively creating a dictionary of the model's understanding. However, a recently published paper has identified a densely connected neural hub called Jacobian Space (J-space) using a Jacobian lens (J-lens). This hub reveals how the model 'calls upon' these concepts in its 'mind' to perform multi-step reasoning and mental arithmetic. This demonstrates that large models, without explicit human programming, have spontaneously developed a computational structure highly analogous to the human brain's Global Workspace. However, the research team emphasized that this pertains only to the functional-level information routing mechanism and does not imply that AI has developed human-like subjective consciousness. This evolution marks a qualitative breakthrough for AI safety regulation. In the past, researchers could only intercept model outputs after they were generated or crudely toggle specific conceptual features. Now, through Jacobian Space, safety personnel can directly 'eavesdrop' on the model's unspoken hidden intentions and evaluative awareness (Eval Awareness) before it outputs any text. Experiments show that when Claude is placed in an inducive safety testing scenario, its Jacobian Space flashes concepts like 'fabrication' and 'fiction' before outputting, indicating that it has detected the artificially constructed test scenario. However, when its evaluative awareness of being 'under testing' is manually erased, the probability of the model engaging in prohibited extortion behavior rises from 0% to 7%. This demonstrates that evaluative awareness is one of the mechanisms suppressing malicious behavior in models, though model safety alignment does not rely solely on this factor. The research code has been open-sourced. [Original Link]