Activation steering offers a lightweight way to control large language model behavior without full fine-tuning. However, existing methods often struggle to balance efficiency with expressiveness, particularly when relying on fixed steering vectors.
The new paper, posted on arXiv, introduces WAQS (Weight Absorption for Quadratic Probing and Affine Steering). The method appears to combine weight absorption with quadratic probing and affine steering to address this trade-off, though the abstract does not provide experimental details.
As the abstract is limited, the effectiveness of WAQS remains to be evaluated. The work is relevant for researchers seeking more flexible and efficient approaches to steering LLM outputs.