Reproduce “Symbolic Working Memory Enhances Language Models for Complex Rule Application” 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
Siyuan Wang, Zhongyu Wei, Yejin Choi, Xiang Ren
A neurosymbolic framework augments Large Language Models with an external symbolic working memory, significantly improving their multi-step deductive reasoning capabilities and robustness, particularly when applying complex rules from non-sequential or noisy inputs. This approach consistently outperforms existing methods across diverse reasoning tasks and different LLM backbones.
Key insights
- Decomposing multi-step reasoning into distinct symbolic rule grounding and LLM-based rule implementation stages significantly enhances LLM performance and robustness in complex deductive tasks.
- Augmenting LLMs with an explicit, structured symbolic working memory addresses their limitations in tracking dynamic information and handling non-sequential or noisy rule inputs more effectively than purely neural approaches.
- The synergistic combination of symbolic precision and neural flexibility offers a promising paradigm for designing more reliable and interpretable AI systems for multi-step logical reasoning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.