Meta Chain-of-Thought: System 2 Reasoning in LLMs (January 2025)

"Towards System 2 Reasoning in LLMs: Learning How to Think With Meta Chain-of-Thought" proposed training models to generate their own reasoning strategies rather than following fixed patterns, inspired by the dual-process theory of human cognition (fast System 1 vs. deliberate System 2 thinking).

Standard chain-of-thought prompting asks models to show their reasoning steps, but the model uses the same approach for every problem. Meta Chain-of-Thought trains the model to first select a reasoning strategy (should I break this into sub-problems? Use analogical reasoning? Work backwards?) and then execute that strategy.

The key insight: models can learn to reason about reasoning. Instead of always doing step-by-step arithmetic, the model might recognize that a geometry problem is better solved by visualization, or that a logic puzzle benefits from working backwards from the answer.

This paper represents the push toward "System 2" reasoning in AI - deliberate, strategic thinking rather than fast pattern matching. It is part of a broader 2025 trend: reasoning models (OpenAI o1, DeepSeek-R1, Claude's extended thinking) that trade compute at inference time for better accuracy.

The practical implication: models that can choose their reasoning strategy perform better on diverse tasks because they apply the right tool for each problem rather than using one approach for everything.

Publication: January 2025. Key contribution: Dynamic reasoning strategy selection for LLMs. Related work: OpenAI o1, DeepSeek-R1, Gemini Deep Think.

Sebastian Raschka: LLM Research Papers 2025 - https://magazine.sebastianraschka.com/p/llm-research-papers-2025-list-one

Towards Data Science: AI Papers 2025 - https://towardsdatascience.com/ai-papers-to-read-in-2025/