Reproduce “Controllable Text Generation with Neurally-Decomposed Oracle” 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
Tao Meng, Sidi Lu, Nanyun Peng, Kai-Wei Chang
A framework called Neurally-Decomposed Oracle (NADO) provides precise control over pre-trained auto-regressive language models by leveraging sequence-level black-box oracles, achieving strong performance in lexical constraint adherence and formality control in machine translation while maintaining generation quality.
Key insights
- Controllable generation can be framed as a posterior regularization problem, allowing for a principled closed-form solution to re-weight token probabilities.
- A sequence-level boolean oracle, which is often easier to define, can be neurally decomposed into effective token-level guidance by predicting the probability of future success.
- Training the auxiliary NADO model on data sampled directly from the base language model and labeled by the black-box oracle avoids the need for external labeled data and mitigates distributional discrepancy.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.