Prompt Compression and Contrastive Conditioning for Controllability and Toxicity Reduction in Language Models
This addresses controllability and toxicity reduction in language models, representing an incremental improvement in decode-time algorithms.
The researchers tackled the problem of controlling language model outputs by compressing prompts, showing that compressed prompts retain substantial information about the original prompt and can guide generation toward desirable text and away from toxicity. They demonstrated that some complex prompts can be effectively compressed into a single token while maintaining compositionality for independent control aspects.
We explore the idea of compressing the prompts used to condition language models, and show that compressed prompts can retain a substantive amount of information about the original prompt. For severely compressed prompts, while fine-grained information is lost, abstract information and general sentiments can be retained with surprisingly few parameters, which can be useful in the context of decode-time algorithms for controllability and toxicity reduction. We explore contrastive conditioning to steer language model generation towards desirable text and away from undesirable text, and find that some complex prompts can be effectively compressed into a single token to guide generation. We also show that compressed prompts are largely compositional, and can be constructed such that they can be used to control independent aspects of generated text.