4.6 KiB
4.6 KiB
Reasoning Techniques Deep Dive
Chain of Thought (CoT) Variants
Zero-Shot CoT
[Task description]
Think step by step before answering.
Use when: Quick reasoning, no examples available Effectiveness: +40-60% on reasoning tasks
Few-Shot CoT
Example 1:
Q: [Question]
A: Let me think through this...
Step 1: [Reasoning]
Step 2: [Reasoning]
Therefore: [Answer]
Example 2:
Q: [Question]
A: Breaking this down...
First: [Reasoning]
Next: [Reasoning]
So: [Answer]
Now solve:
Q: [Your question]
Use when: Complex reasoning, pattern demonstration needed Effectiveness: +50-80% on complex tasks
Auto-CoT
Let me approach this systematically:
1. Identify the key elements
2. Analyze relationships
3. Apply relevant principles
4. Draw conclusions
5. Verify my reasoning
Use when: General problem-solving, exploratory reasoning
Tree of Thoughts (ToT)
Implementation Pattern
Problem: [Complex problem]
Generate 3 different approaches:
Approach A:
- Method: [Description]
- Reasoning: [Why this might work]
- Potential issues: [Risks]
- Confidence: [1-10]
Approach B:
- Method: [Description]
- Reasoning: [Why this might work]
- Potential issues: [Risks]
- Confidence: [1-10]
Approach C:
- Method: [Description]
- Reasoning: [Why this might work]
- Potential issues: [Risks]
- Confidence: [1-10]
Evaluate branches:
- Which has highest success probability?
- Which has fewest risks?
- Which is most feasible?
Selected approach: [Best option with justification]
Execution: [Step-by-step implementation]
Use when: Strategic decisions, multiple valid paths, high-stakes problems
Self-Consistency
Process
- Generate 5 responses at temp=0.7
- Extract final answers from each
- Take majority vote
- Report confidence = agreement %
Implementation
Run this prompt 5 times (or use n=5 parameter):
[Your reasoning task]
Think step by step and provide final answer.
Aggregate: If 4/5 agree = high confidence
If 3/5 agree = medium confidence
If split = low confidence, needs review
Use when: Math, logic, factual questions with verifiable answers
ReAct (Reasoning + Acting)
Full Pattern
Task: [Goal to achieve]
Thought 1: I need to understand the current situation.
Action 1: [Observation or tool use]
Observation 1: [Result from action]
Thought 2: Based on this, I should [next logical step].
Action 2: [Next action]
Observation 2: [Result]
Thought 3: Now I can see that [insight].
Action 3: [Verification or next step]
Observation 3: [Result]
Thought 4: I have enough information to conclude.
Final Answer: [Conclusion with reasoning]
Use when: Tool-augmented reasoning, research tasks, multi-step analysis
Least-to-Most Prompting
Structure
Complex problem: [Full problem statement]
Step 1: Decomposition
Break this into simpler subproblems, ordered from easiest to hardest:
1. [Simplest subproblem]
2. [Next subproblem, may depend on 1]
3. [Harder subproblem, may depend on 1,2]
4. [Final subproblem requiring all above]
Step 2: Sequential Solution
Subproblem 1: [Solution]
Using result from 1, Subproblem 2: [Solution]
Using results from 1,2, Subproblem 3: [Solution]
Using all results, Subproblem 4: [Solution]
Final integrated answer: [Complete solution]
Use when: Mathematical word problems, multi-step procedures, compositional tasks
Decomposed Prompting (DECOMP)
Pattern
Task: [Complex task]
Required capabilities:
- [Capability 1]: Use [specialized prompt/tool]
- [Capability 2]: Use [specialized prompt/tool]
- [Capability 3]: Use [specialized prompt/tool]
Orchestration:
1. Call [Capability 1] with [input] → get [output1]
2. Call [Capability 2] with [output1] → get [output2]
3. Call [Capability 3] with [output2] → get [final output]
Integrate results: [Final response]
Use when: Tasks requiring diverse expertise, specialized sub-tasks
Constitutional AI Reasoning
Self-Critique Pattern
Initial response: [Generated content]
Critique against principles:
- Is it helpful? [Yes/No + reasoning]
- Is it harmless? [Yes/No + reasoning]
- Is it honest? [Yes/No + reasoning]
If any No:
Revised response that addresses [specific issues]:
[Improved content]
Choosing the Right Technique
| Task Type | Best Technique |
|---|---|
| Simple reasoning | Zero-shot CoT |
| Complex multi-step | Few-shot CoT |
| Strategic decisions | Tree of Thoughts |
| Factual verification | Self-Consistency |
| Tool-using tasks | ReAct |
| Word problems | Least-to-Most |
| Specialized sub-tasks | DECOMP |
| Safety-critical | Constitutional AI |