Reproduce “COLD Decoding: Energy-based Constrained Text Generation with Langevin Dynamics” 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
Lianhui Qin, Sean Welleck, Daniel Khashabi, Yejin Choi
COLD decoding introduces an energy-based framework for constrained text generation that leverages Langevin dynamics on a continuous representation of text. This approach enables off-the-shelf large language models to produce fluent text satisfying diverse, user-defined constraints without requiring task-specific fine-tuning, demonstrating strong performance across tasks like abductive reasoning and counterfactual rewriting.
Key insights
- Gradient-based Langevin dynamics can be effectively adapted to discrete text generation via continuous relaxation, enabling efficient sampling from energy-based models for language.
- A unified energy function can flexibly combine various hard lexical and soft contextual/semantic constraints, offering a versatile framework for controlling text generation.
- Effective constrained text generation can be achieved without fine-tuning large, off-the-shelf language models, providing a practical solution for leveraging powerful but inaccessible or computationally expensive models.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.