Reproduce “Improving Coherence and Consistency in Neural Sequence Models with Dual-System, Neuro-Symbolic Reasoning” 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
Maxwell Nye, Michael Henry Tessler, Joshua B. Tenenbaum, Brenden M. Lake
The authors present a dual-system approach that filters neural story continuations through a symbolic world model. A parser converts proposals into facts, and a constraint checker rejects incompatible ones before resampling. Participants preferred its continuations in 84% (95% bootstrap CI 0.77–0.89). On one low-data gSCAN split, exact-match accuracy rose from 68.1% to 81.3%, while novel-direction accuracy remained near zero.
Key insights
- Integrating explicit symbolic reasoning as a 'System 2' module effectively corrects logical inconsistencies produced by 'System 1-like' neural generators.
- Few-shot parsing with large language models can serve as a lightweight and training-free bridge to translate neural text into symbolic representations for logical evaluation.
- The dual-system approach, inspired by cognitive science's dual-process theories, demonstrates that combining intuitive neural processing with deliberative symbolic reasoning enhances AI model coherence and robustness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.