Reasoning in Natural Language
summary
Current large LMs show fluent but brittle behavior: they hallucinate facts, fail spatial and scheduling puzzles, and miss simple logical set and counting inferences. Research group builds an extraction pipeline that yields 2.6 million sentence pairs to train a decomposition model called DECcompose-T5 and shows improved QA on compositional queries. Other work prototypes Dialogue2APIs to let LMs plan, call external tools and backtrack. A multi-perspective provenance system is released as a website and influences commercial search experiments.
AI-written summary · about
chapters (8)
- 0:00Intro
- 0:45AI in the news
- 2:15So where are we?
- 9:56Spatial Reasoning
- 20:17Key Challenges to NLP Researchers
- 25:57DIalog2APIs
- 34:23Comparable Texts Extraction
- 43:10Conclusion
related talks
01h 31mNatural Language Understanding: Foundations and State-of-the-Art
01h 00mHuman and machine inductive biases for compositional linguistic generalization
36mEnhanced and Efficient Reasoning in Large Language Models
49mThe Building Blocks of Today’s and Tomorrow’s Language Models
27mExploring the Mechanisms of LLM Reasoning Capabilities
39mUnlocking reasoning in Large language models
55mLessons Learned From Implementing Generative AI Solutions
16mBeyond LLMs - The Next Frontier of Machine Reasoning