Reproduce “Reward-Augmented Decoding: Efficient Controlled Text Generation With a Unidirectional Reward Model” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Haikang Deng, Colin Raffel
RAD steers a large language model during generation without retraining it. It restricts choices to top-k tokens, scores continuations with a smaller left-to-right reward model, and reuses cached prefix representations. In GPT-2 Large detoxification, it reports Average Max Toxicity of 0.148, below Rectification's 0.180 and Quark's 0.196; this finding is limited to that tested task.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.