ReAct vs Chain-of-Thought

1Question: If I have 3 apples and buy 2 more, then give half
2to my friend, how many do I have left?
3
4Chain-of-Thought Response:
5
6Let me work through this step by step:
7
81. I start with 3 apples
92. I buy 2 more apples: 3 + 2 = 5 apples
103. I give half to my friend: 5 / 2 = 2.5 apples
114. Since I can't give half an apple, I round down: 2 apples
12
13Wait, let me reconsider. "Half" likely means half of the total.
145 / 2 = 2.5, and I give away 2.5, keeping 2.5.
15
16If we're talking about whole apples, I'd give 2 and keep 3,
17or give 3 and keep 2.
18
19Answer: I have 2 or 3 apples left, depending on how we
20handle the half.

1# Zero-shot CoT
2zero_shot_prompt = """
3Question: {question}
4
5Let's think step by step.
6"""
7
8# Few-shot CoT
9few_shot_prompt = """
10Question: What is 15% of 80?
11
12Let's think step by step:
131. 15% means 15/100 = 0.15
142. 0.15 × 80 = 12
15
16Answer: 12
17
18Question: {question}
19
20Let's think step by step:
21"""
22
23# Self-consistency CoT (generate multiple, take majority)
24def self_consistent_cot(question: str, n_samples: int = 5) -> str:
25    answers = []
26    for _ in range(n_samples):
27        response = llm.generate(f"{question}\nLet's think step by step:")
28        answer = extract_final_answer(response)
29        answers.append(answer)
30
31    # Return most common answer
32    from collections import Counter
33    return Counter(answers).most_common(1)[0][0]

1Task: "What is the current stock price of Apple?"
2
3CHAIN-OF-THOUGHT:
4"Let me think about this...
5Apple is a major tech company traded as AAPL.
6Stock prices fluctuate constantly.
7Based on recent trends, it might be around $180-190.
8
9Answer: Approximately $185 (note: this is an estimate)"
10
11Problem: The answer is a guess from training data!
12
13---
14
15REACT:
16Thought: I need to find the current stock price.
17         This requires up-to-date information.
18Action: search(query="AAPL stock price today")
19Observation: AAPL is trading at $193.42, up 1.2% today.
20
21Thought: I have the current price from a reliable source.
22Action: finish(answer="Apple (AAPL) is currently at $193.42")
23
24Benefit: Actual current information, not a guess.

1CHAIN-OF-THOUGHT EXCELS:
2- Mathematical reasoning
3- Logical deduction
4- Problems with complete information
5- Analysis of given text
6
7Example: "Given that all dogs are mammals and all mammals
8         are warm-blooded, are dogs warm-blooded?"
9
10CoT: "Let me reason through this:
111. All dogs are mammals (given)
122. All mammals are warm-blooded (given)
133. Therefore, since dogs are mammals, dogs are warm-blooded.
14
15Answer: Yes, dogs are warm-blooded."
16
17(No external information needed - pure deduction)
18
19---
20
21REACT EXCELS:
22- Tasks requiring current information
23- Multi-step research
24- Code execution and verification
25- Interacting with systems
26
27Example: "Is the website example.com currently online?"
28
29ReAct:
30Thought: I need to check if the website is responding.
31Action: check_website(url="https://example.com")
32Observation: Status 200 OK, response time 45ms
33
34Answer: Yes, example.com is online and responding normally.
35
36(Required actual verification)

1class HybridAgent:
2    """Combines CoT for reasoning, ReAct for action."""
3
4    def __init__(self, llm, tools):
5        self.llm = llm
6        self.tools = tools
7
8    def solve(self, task: str) -> str:
9        # Step 1: CoT analysis to understand the task
10        analysis = self.analyze_with_cot(task)
11
12        # Step 2: Determine if external action needed
13        needs_action = self.needs_external_action(analysis)
14
15        if not needs_action:
16            # Pure reasoning is sufficient
17            return self.solve_with_cot(task)
18        else:
19            # Use ReAct for grounded execution
20            return self.solve_with_react(task, analysis)
21
22    def analyze_with_cot(self, task: str) -> str:
23        """Use CoT to analyze the task."""
24        prompt = f"""
25Analyze this task step by step:
26
27Task: {task}
28
29Consider:
301. What type of problem is this?
312. What information do I need?
323. Is all needed information available, or do I need to gather more?
334. What's my approach?
34
35Analysis:
36"""
37        return self.llm.generate(prompt)
38
39    def needs_external_action(self, analysis: str) -> bool:
40        """Determine if external action is needed."""
41        external_indicators = [
42            "need to search",
43            "need to check",
44            "need to verify",
45            "current information",
46            "real-time",
47            "execute",
48            "run",
49            "file",
50        ]
51
52        analysis_lower = analysis.lower()
53        return any(ind in analysis_lower for ind in external_indicators)
54
55    def solve_with_cot(self, task: str) -> str:
56        """Solve using pure Chain-of-Thought."""
57        prompt = f"""
58{task}
59
60Let me think through this step by step:
61"""
62        response = self.llm.generate(prompt)
63        return self.extract_answer(response)
64
65    def solve_with_react(self, task: str, analysis: str) -> str:
66        """Solve using ReAct with tools."""
67        # Initialize ReAct agent
68        react = ReActAgent(self.llm, self.tools)
69        # Include analysis as context
70        enhanced_task = f"{task}\n\nInitial analysis:\n{analysis}"
71        return react.run(enhanced_task)

1Task: "Calculate the total cost of items in cart.json and
2apply a 15% discount if over $100"
3
4Thought: I need to first read the cart to see the items and prices.
5Action: read_file(path="cart.json")
6Observation: {"items": [{"name": "Book", "price": 29.99},
7             {"name": "Pen", "price": 4.99},
8             {"name": "Notebook", "price": 12.99}]}
9
10Thought: Let me calculate the total using chain-of-thought reasoning:
11- Book: $29.99
12- Pen: $4.99
13- Notebook: $12.99
14- Total: 29.99 + 4.99 + 12.99 = $47.97
15
16Since $47.97 < $100, no discount applies.
17
18Action: finish(answer="Total: $47.97. No discount applied (under $100)")
19
20The agent used CoT reasoning WITHIN the ReAct thought phase!