arXiv · 2023 · Preprint
Turner et al. · → Paper · Demo: ✗ · Code: ✓
Introduces activation engineering and the ActAdd technique, which steers language model outputs at inference time by adding a difference vector computed from a contrastive prompt pair directly to the residual stream of a frozen transformer, achieving state-of-the-art toxicity reduction and sentiment control without any gradient-based optimization.
Citation Stub
This paper is not a speech generation paper but is cited by the corpus. See Context in Speech Generation below for why it is relevant.
Context in Speech Generation
ActAdd demonstrates that the internal activation space of a large language model encodes high-level semantic and stylistic attributes as approximately linear directions, and that adding a lightweight steering vector to the residual stream at a single layer during inference is sufficient to redirect generation toward a target property. For spoken conversational agents and instruction-conditioned speech systems that use a language model backbone, this technique provides a mechanism for inference-time behavioral control that requires no weight updates and as few as two contrastive examples. The method preserves off-target task performance (as measured on ConceptNet factual recall), suggesting that steering vectors can be applied without degrading the model’s general capabilities. Speech generation systems that expose intermediate activations of their language model components can in principle adopt activation addition as a lightweight alternative to prompting or fine-tuning for attribute control.
Wiki Connections
- spoken-language-model — ActAdd provides an inference-time steering mechanism applicable to the language model backbones that underpin speech LMs and spoken conversational agents.
- instruction-conditioned-tts — the activation engineering paradigm offers an alternative to natural-language instruction conditioning, enabling attribute control without occupying context window space or requiring fine-tuning.