Decomposing how prompting steers behavior

In three linesStudy of representational geometry to understand how prompts reshape behavior in LLMs and VLMs. Nested decomposition framework testing translation, rigid transformation, scaling, affine and nonlinear maps on 3 LLMs, 3 VLMs and 6 datasets. Finding: cross-dimensional linear mixing (affine transformation) is the key mechanism for representational reorganization toward task structure.

Summary generated by Claude — human-verified