ReAct (Reason + Act)
About
ReAct (Reason + Act) is a prompting technique where the model alternates between reasoning and taking actions to solve a problem.
Instead of only thinking internally (like Chain-of-Thought), the model:
Reasons about the problem
Decides an action to take
Uses external information or tools (if available)
Observes the result
Continues reasoning based on new information
Core idea:
Combine reasoning with interaction to improve decision-making.
ReAct introduces a loop:
Reason → Act → Observe → Repeat → Final Answer
This makes the model behave more like an agent, not just a responder.
How ReAct Works (Model Behavior Perspective)
LLMs are trained on sequences that include:
Problem solving
Question answering
Tool usage patterns (search, lookup, API calls)
Multi-step reasoning
ReAct structures this into a repeatable loop:
Thought (Reasoning) Model analyzes the current situation
Action Model decides what to do next (e.g., query, lookup, calculate)
Observation Model receives new information (from context or tool output)
Next Thought Model updates reasoning using new data
This continues until the model reaches a final answer.
Example pattern:
Thought: I need more information Action: Search for X Observation: Result Y Thought: Now I can conclude… Final Answer: Z
This structure prevents premature conclusions.
Strengths and Ideal Use Cases
1. Handles Incomplete or Dynamic Information
ReAct is useful when:
All data is not available upfront
Information must be retrieved step by step
Decisions depend on intermediate findings
Examples:
Knowledge retrieval systems
API-based workflows
Debugging unknown errors
Investigation tasks
2. Reduces Hallucination
Instead of guessing, the model:
Actively retrieves or uses information
Bases reasoning on observations
Updates conclusions dynamically
This reduces reliance on internal assumptions.
3. Enables Tool Integration
ReAct is the foundation for:
AI agents
Tool-calling systems
API integrations
Retrieval pipelines
The “Act” step can represent:
Database queries
API calls
Search operations
Code execution
4. Improves Multi-Step Decision Making
For complex workflows:
Each step depends on previous output
Decisions must adapt dynamically
ReAct ensures:
Continuous feedback loop
Incremental reasoning
Better alignment with real-world processes
Limitations and Practical Considerations
1. Increased Complexity
ReAct introduces:
Multi-step interaction
State management
Intermediate outputs
This makes prompts and systems more complex to design and maintain.
2. Higher Cost and Latency
Each cycle (Reason → Act → Observe):
Consumes tokens
May involve external calls
Increases response time
Not suitable for simple or real-time low-latency tasks.
3. Requires Structured Environment
For full effectiveness, ReAct often needs:
Defined tools or APIs
Controlled execution environment
Clear action definitions
Without this, “actions” become simulated rather than real.
4. Risk of Incorrect Actions
If reasoning is flawed:
Model may choose wrong actions
Incorrect observations may propagate errors
Proper constraints and validation are important.
Sample Prompts
Without ReAct (Direct Reasoning)
Issue:
Model may guess or hallucinate exchange rate
No real-time validation
No external interaction
With ReAct (Reason + Act Pattern)
Expected behavior:
Model identifies need for exchange rate
Simulates or triggers retrieval
Uses retrieved value
Computes final answer
Last updated