Reproduce “Learning Reasoning Strategies in End-to-End Differentiable Proving” 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
Pasquale Minervini, Sebastian Riedel, Pontus Stenetorp, Edward Grefenstette, Tim Rocktäschel
The UCL Centre for Artificial Intelligence and Facebook AI Research developed Conditional Theorem Provers (CTPs), which enhance neuro-symbolic reasoning by learning to dynamically select relevant rules for a given goal, overcoming the scalability limitations of prior Neural Theorem Provers. These CTPs achieved superior systematic generalization on relational reasoning tasks and competitive link prediction performance on knowledge graphs, while providing interpretable logical rules.
Key insights
- Adaptive, goal-conditioned rule selection can dramatically reduce the computational burden of differentiable theorem proving, making it scalable for complex reasoning problems.
- By learning to reformulate goals into relevant subgoals using continuous embeddings, the model can discover explicit and interpretable logical reasoning strategies directly from data.
- Integrating a differentiable rule selection mechanism allows for effective joint training of both symbolic rule application and neural representation learning within an end-to-end framework.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.