Modèles de raisonnement
Définition
Les patterns de raisonnement sont des moyens structurés pour susciter ou organiser le raisonnement du modèle: chain-of-thought (step-by-step), tree-of-thoughts (explore branches), ReAct (reason + act), and RDD (récupération-décision-conception), among others. Using a clear pattern improves reliability (more consistent raisonnement) and debuggability (you can inspect steps or actions).
Ils sont used in prompt engineering (par ex. CoT) and inside agents (par ex. ReAct, RDD). Choosing a pattern depends on the task: CoT for math/raisonnement, ReAct for tool use, ToT for search/planning, RDD for spec compliance.
Comment ça fonctionne
You feed input (question, task) into a pattern: the pattern constrains how the model reasons or acts (par ex. “think étape par étape”, or thought–action–observation loops). The model produces an output (answer, action sequence). Prompts or system conception encourage the model to show raisonnement (par ex. “Think étape par étape”) or to interleave thought and action. Patterns can be combined (par ex. CoT inside an agent loop). See the linked pages for each pattern’s details.
Cas d'utilisation
Different patterns suit different needs: CoT for stepwise raisonnement, ReAct for tool use, ToT for search and planning.
- CoT: math, logic, and multi-step raisonnement tasks
- ReAct: tool-using agents that reason before each action
- ToT: search and planning over multiple solution branches
Documentation externe
- Chain-of-Thought Prompting (Wei et al.) — CoT paper
- ReAct: Synergizing Reasoning and Acting (Yao et al.) — ReAct paper
- Tree of Thoughts (Yao et al.) — ToT paper