Independent interpretability researcher (quantitative-biology and statistics background) studying how transformers organize next-token computation in the residual stream, and the verification gap in AI-assisted research. Open to collaboration and feedback.
Nelson Guda
Karma: 1
Super interesting work. The first thing that jumped out at me was that the examples of <think> text that you give start with “The user...”, and I was wondering if all internal context tends to style this way in the models you looked at. In the destyled example, the text starts is only slightly different “The user’s...”. This makes me curious about what you found in the destyling experiments with regard to different semantic structures that are more or less close to this.
Thanks for sharing.