LLM Metacognition: Understanding and leveraging "Thinking about Thinking"
summary
Claim: contemporary LLMs exhibit metacognitive labels that let them name and cluster procedural skills and use those labels to improve other models. Evidence: ~100 Wikipedia-derived skills are combined by GPT-4 into k-skill texts (small models fail; GPT-4 solves k≈3–5), and a theory predicts each 10× scale roughly doubles compositional reach. Tools and numbers: fine-tuning on GPT-4–generated k≤3 examples generalizes to k=4,5; skill-guided in-context selection boosts accuracy up to 5%; human vetting leaves 56% of synthetic MATH questions unchanged and 7% heavily edited.
AI-written summary · about
chapters (11)
- 0:00Introduction
- 1:15LLMs as parrots and metacognition
- 9:09Compositional capability and scaling
- 16:15Can small LLMs learn skills?
- 20:07LLM metacognition concept
- 23:26Extracting and validating skills
- 33:26Instruction skills and evaluation
- 35:34MATH Square dataset
- 44:09Instruction tuning with metacognition
- 48:56Theory of skill emergence
- 59:16Q&A
related talks
01h 03mLLM skills and meta-cognition: scaffolding for new forms of learning?
01h 18mThe Enigma of LLMs: on Creativity, Compositionality, Pluralism, and Paradoxes
19mLLMs reasoning and agentic capabilities over time
26mGrappling with the Bitter Lesson
17mPsychology of AI
27mExploring the Mechanisms of LLM Reasoning Capabilities
25mWhy LLMs Still Fail: Reasoning, Robustness, and Safety
44mUnderstanding the abilities of AI systems: Memorization, generalization, and points in between