Reasoning & Interpretability (2013-2023)

Methods for improving model reasoning and understanding model decisions.

Chain-of-Thought & Multi-Path Reasoning

• Chain-of-Thought Prompting Elicits Reasoning in Large Language Models - CoT prompting
• Tree of Thoughts: Deliberate Problem Solving with LLMs - Multi-path reasoning
• Self-Consistency Improves Chain of Thought Reasoning - Ensembling reasoning paths

Interpretability & Explanation

• Attention Is Not Explanation - Attention visualization limits
• The Saliency MAP Shows What Regular Visualization Cannot - Visual explanations

---

Key Insight: Reasoning techniques and interpretability methods are essential for understanding and improving LLM behavior.