WRSWorld Research StateContribute
WORLD RESEARCH STATE

Turner and colleagues: prompt-pair activation addition

Contribution grade

Language-model representation control · Human-designed model analysis or safeguard method

Reference score50,000Provisional editorial estimate of this exact contribution; its grade is assessed separately.

What changed

Used activation differences elicited by contrasting prompts as inference-time steering vectors, showing selected changes to generated text without gradient-based fitting of the steering direction.

Scientific scope

Used activation differences elicited by contrasting prompts as inference-time steering vectors, showing selected changes to generated text without gradient-based fitting of the steering direction. No optimization of the vector does not mean no parameter selection or computational cost: layer and strength choices still matter and internal activation access is required. The demonstrated controls are not universal concept disentanglement.

Evidence

The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. No optimization of the vector does not mean no parameter selection or computational cost: layer and strength choices still matter and internal activation access is required. The demonstrated controls are not universal concept disentanglement.

Contributors

Provisional WRS assessment, edition 2026-10-23.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction