Tag
A researcher presents evidence that strong target text can induce a measurable latent-state shift in Gemma 3 12B IT before final output, distinct from lexical or content overlaps, and discusses implications for AI safety beyond output-only evaluation.