New framework maps 'Learning by Thinking' in LLMs

A new paper proposes a framework to understand how large language models (LLMs) exhibit 'Learning by Thinking' (LbT) mechanisms, similar to human cognitive processes. The 'Cognitive LbT Framework' identifies four key mechanisms: Reflective Reasoning, Synthetic Memory Formation, Internal Simulation, and Conceptual Self-Organization. The paper analyzes how current methods like Chain-of-Thought prompting and Tree-of-Thoughts search functionally approximate these mechanisms, leading to performance improvements on tasks without new labeled data, such as a reported 33% gain on Super-NaturalInstructions. However, these LLM LbT processes also result in characteristic failures like hallucination diffusion and high computational costs, stemming from structural asymmetries with human cognition. AI

IMPACT Provides a structured way to analyze and potentially improve LLM reasoning and learning capabilities.

RANK_REASON The item is an academic paper proposing a new framework for understanding LLM capabilities. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• chain-of-thought prompting
• Cognitive LbT Framework
• Conceptual Self-Organization
• Internal simulation of expected sensory experiences before movements get started.
• large-language models
• Learning by Thinking
• Reflective Reasoning
• Self-Instruct
• self-reflection
• self-taught reasoning
• Synthetic Memory Formation
• Tree of Thoughts
• Y. Wang

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →