Out-of-Distribution Generalization as Reasoning: Are LLMs Competitive?
summary
Proposes a concrete alternative to LLM-first reasoning: learn boolean relational recognizers and chain them with probabilistic soundness so if components are 90% supported a chained conclusion is provably ~80% supported; uses PAC-style bounds and union bounds, limits arity to keep runtime polynomial, and tests a prototype on Wall Street Journal text where chaining outperforms plain ML. Describes tokenized relational features, obscured examples, and an implemented robust logic pipeline in a Wall Street Journal experiment that demonstrates feasibility.
AI-written summary · about
chapters (16)
- 0:00Introduction
- 1:03Speaker introduction
- 2:09Human cognition mystery
- 4:17Why large language models may not solve reasoning directly
- 5:44Robust logic introduction
- 12:06Robust logic features
- 16:16Knowledge representation
- 22:15Reasoning in AI
- 23:41Darwin camera example
- 31:23Modular learnability limits
- 36:31Polynomial-time reasoning
- 46:02Trustworthiness vs accuracy
- 51:31Reasoning benchmarks
- 58:09LLMs vs robust logic
- 1:00:19Feature extraction + reasoning
- 1:01:44Closing remarks
related talks
27mExploring the Mechanisms of LLM Reasoning Capabilities
36mEnhanced and Efficient Reasoning in Large Language Models
29mThe Evolution of LLMs: From Imitation to Reasoning
28mSymbolic Reasoning in telecom oriented LLMs
31mBeyond LLMs and Maturing RAG use cases
16mBeyond LLMs - The Next Frontier of Machine Reasoning
22mPractical Strategies for Implementing Domain-Specific LLMs
25mWhy LLMs Still Fail: Reasoning, Robustness, and Safety