Reproduce “DECIDER: A Dual-System Rule-Controllable Decoding Framework for Language Generation” 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
Chen Xu, Tian Lan, Yu Ji, Changlong Yu, Wei Wang, Jun Gao, Qunxi Dong, Kun Qian and 3 more
DECIDER introduces a dual-system framework, inspired by cognitive dual-process theory, that integrates a pre-trained language model with a First-Order Logic reasoner for rule-controllable text generation. This approach generates more natural, coherent, and informative text, outperforming baselines on CommonGen and PersonaChat tasks while consistently enhancing various LLMs.
Key insights
- The cognitive dual-process theory, separating intuitive neural generation from symbolic logical reasoning, offers an effective paradigm for enhancing controllable language generation.
- Guiding LLMs with high-level, programmable First-Order Logic rules provides more nuanced and intelligent control over text semantics and style compared to encouraging specific target words.
- Employing soft logic (continuous truth values for predicates) is more effective than hard (binary) logic for achieving nuanced control, and injecting logical reasoning signals at multiple points in the decoder (attention and prediction distributions) is crucial.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.