Enhanced and Efficient Reasoning in Large Language Models
summary
Proposes a two-stage pipeline that first recodes text into a Unary Relational Integracode (expand each entity into ~4 local Boolean/embedding slots) and then trains a standard classifier to predict relations, enabling robust logic style probabilistic chaining across learned rules. Claims a reduction to a k-DNF mapping (k set by rule arity) that makes core relational-rule learning polynomial-time in rule complexity under simple-scene heuristics. Demonstrated on a small Wall Street Journal experiment (246 target words) with measurable prediction gains for many words.
AI-written summary · about
chapters (8)
- 0:00Introduction
- 1:27Trust gap in LLMs for life advice
- 5:27Principled reasoning for LLMs
- 9:04Robust logic principles
- 14:14Early robust logic experiments
- 20:45Token representations for reasoning
- 28:04Theoretical foundations of robust logic
- 31:35Q&A
related talks
01h 02mOut-of-Distribution Generalization as Reasoning: Are LLMs Competitive?
45mReasoning in Natural Language
39mUnlocking reasoning in Large language models
29mThe Evolution of LLMs: From Imitation to Reasoning
50mKnowledge is embedded in language neural networks but can they reason?
42mRDF and the future of LLMs
32mScaling Test-time Inference Compute & Rise of Reasoning Models
48mProbabilistic Reasoning by First-Order Model Counting