Chain of Thought
Prompting a model to reason step by step before answering.
Aperçu
Chain of thought asks a model to produce intermediate reasoning before its conclusion, which measurably improves accuracy on problems that need several steps. It solves the problem that a model answering immediately has to reach the answer in one forward pass, with no room to work. It exists because it costs only tokens and needs no training, and it is the clearest case where how a model is asked changes what it can do.
Graphe des technologies
Comment Chain of Thought se rattache au reste de la pile.
Utilise et dépend de
- Self-Consistency technique Prompting Self-Consistency based on Chain of Thought High ✓ source
- Tree of Thoughts technique Prompting Tree of Thoughts based on Chain of Thought High ✓ source
Utilisé par
Appliqué à
- Prompt Engineering technique Prompting Chain of Thought used for Prompt Engineering High ✓ source
- Reinforcement Learning concept Algorithms Chain of Thought used for Reinforcement Learning Medium ✓ source
- GSM8K benchmark Benchmarks Chain of Thought solves GSM8K High ✓ source