Use when tackling complex reasoning tasks requiring step-by-step logic, multi-step arithmetic, commonsense reasoning, symbolic manipulation, or problems where simple prompting fails - provides comprehensive guide to Chain-of-Thought and related prompting techniques (Zero-shot CoT, Self-Consistency,

Thought Based Reasoning Techniques for LLMs Overview Chain of Thought (CoT) prompting and its variants encourage LLMs to generate intermediate reasoning steps before arriving at a final answer, significantly improving performance on complex reasoning tasks. These techniques transform how models approach problems by making implicit reasoning explicit. Quick Reference Technique When to Use Complexity Accuracy Gain Zero shot CoT Quick reasoning, no examples available Low +20 60% Few shot CoT Have good examples, consistent format needed Medium +30 70% Self Consistency High stakes decisions, need confidence Medium +10 20% over CoT Tree of Thoughts Complex problems requiring exploration High +50 70% on hard tasks Least to Most Multi step problems with subproblems Medium +30 80% ReAct Tasks requiring external information Medium +15 35% PAL Mathematical/computational problems Medium +10 15% Reflexion Iterative improvement, learning from errors High +10 20% Core Techniques 1. Chain of Thought (CoT) Prompting Paper : "Chain of Thought Prompting Elicits Reasoning in Large Language Models" (Wei et al., 2022) Citations : 14,255+ When to Use Multi step arithmetic or math word problems Commonsense reasoning requiring logical deduction Symbolic reasoning tasks When you have good exemplars showing reasoning How It Works Provide few shot examples that include intermediate reasoning steps, not just question answer pairs. The model learns to generate similar step by step reasoning. Prompt Template Strengths Significant accuracy improvements on reasoning tasks Interpretable intermediate steps Works well with large models ( 100B parameters) Limitations Requires crafting good exemplars Less effective on smaller models Can still make calculation errors 2. Zero shot Chain of Thought Paper : "Large Language Models are Zero Shot Reasoners" (Kojima et al., 2022) Citations : 5,985+ When to Use No exemplars available Quick reasoning needed General purpose reasoning across task types Prototyping before creating few shot examples How It Works Simply append "Let's think step by step" (or similar phrase) to the prompt. This triggers the model to generate reasoning steps without any examples. Prompt Template Alternative trigger phrases : "Let's work this out step by step to be sure we have the right answer." "Let's break this down." "Let's approach this systematically." "First, let me understand the problem..." Two Stage Approach (More Robust) Stage 1 Reasoning Extraction : Stage 2 Answer Extraction : Strengths No exemplar crafting required Generalizes across task types Simple to implement Limitations Less effective than few shot CoT Can produce verbose or irrelevant reasoning Sensitive to exact phrasing 3. Self Consistency Paper : "Self Consistency Improves Chain of Thought Reasoning in Language Models" (Wang et al., 2022) Citations : 5,379+ When to Use High stakes decisions requiring confidence Problems with multiple valid reasoning paths When you need to reduce variance in outputs Verification of reasoning correctness How It Works Sample multiple diverse reasoning paths, then select the most consistent answer via majority voting. The intuition: correct answers can be reached through multiple reasoning paths. Prompt Template Implementation Example Strengths Significant accuracy boost over single path CoT Provides confidence measure (agreement level) Task agnostic improvement Limitations Higher computational cost (N times more generations) Requires extractable discrete answers Diminishing returns beyond ~10 20 samples 4. Tree of Thoughts (ToT) Paper : "Tree of Thoughts: Deliberate Problem Solving with Large Language Models" (Yao et al., 2023) Citations : 3,026+ When to Use Complex problems requiring exploration/backtracking Tasks where initial decisions are pivotal Creative problem solving (writing, puzzles) When CoT alone achieves <50% accuracy How It Works Generalize CoT to a tree structure where each node is a "thought" (coherent language unit). Uses search algorithms (BFS/DFS) with self evaluation to explore and select promising reasoning paths. Prompt Template Thought Generation : State Evaluation : BFS/DFS Search : Example: Game of 24 Strengths Dramatically improves performance on hard tasks (4% → 74% on Game of 24) Enables backtracking and exploration Self evaluation catches errors early Limitations Significantly higher computational cost Requires task specific thought decomposition Complex to implement 5. Least to Most Prompting Paper : "Least to Most Prompting Enables Complex Reasoning in Large Language Models" (Zhou et al., 2022) Citations : 1,466+ When to Use Problems harder than your exemplars Compositional generalization tasks Multi step problems with clear subproblems Symbol manipulation and SCAN like tasks How It Works Two stage process: 1. Decomposition : Break complex problem into simpler subproblems 2. Sequential Solving : Solve subproblems in order, using previous answers Prompt Template Stage 1: Decomposition Stage 2: Sequential Solving Strengths Excellent at generalizing to harder problems Works well on compositional tasks Explicit problem decomposition aids interpretability Limitations Requires two stage prompting Decomposition step can fail on novel structures More complex setup than single stage CoT 6. ReAct (Reasoning + Acting) Paper : "ReAct: Synergizing Reasoning and Acting in Language Models" (Yao et al., 2022) Citations : 5,012+ When to Use Tasks requiring external information (search, APIs) Interactive decision making environments Multi hop question answering When pure reasoning leads to hallucination How It Works Interleave reasoning traces ("Thought") with actions ("Action") and observations ("Observation"). Reasoning helps plan actions; actions provide new information for reasoning. Prompt Template Action Types Search[query] Search for information Lookup[keyword] Look up keyword in current context Finish[answer] Return final answer Strengths Reduces hallucination by grounding in external knowledge Interpretable action traces Handles exceptions through adaptive reasoning Limitations Requires integration with external tools More complex orchestration Action space must be defined 7. PAL (Program Aided Language Models) Paper : "PAL: Program aided Language Models" (Gao et al., 2022) Citations : 608+ When to Use Mathematical/arithmetic reasoning Problems requiring precise computation Symbolic manipulation When CoT makes calculation errors How It Works Generate code (typically Python) instead of natural language reasoning. Execute the code to get the answer. The LLM handles decomposition; the interpreter handles computation. Prompt Template Strengths Eliminates arithmetic errors Clear variable naming aids interpretability Leverages code execution for verification Limitations Requires code interpreter Not suitable for non computational reasoning Model must generate syntactically correct code 8. Auto CoT Paper : "Automatic Chain of Thought Prompting in Large Language Models" (Zhang et al., 2022) Citations : 838+ When to Use No manually crafted exemplars available Want to automate few shot CoT setup Scaling CoT to many tasks When zero shot CoT isn't sufficient How It Works 1. Cluster questions by diversity 2. Use Zero shot CoT to generate reasoning chains for representative questions 3. Use these auto generated chains as few shot exemplars Prompt Template Step 1: Generate diverse demonstrations Step 2: Use as few shot exemplars Strengths No manual exemplar creation Diversity sampling improves robustness Matches manual CoT performance Limitations Quality depends on zero shot CoT quality Clustering requires similarity metric Some generated chains contain errors 9. Reflexion Paper : "Reflexion: Language Agents with Verbal Reinforcement Learning" (Shinn et al., 2023) Citations : 2,179+ When to Use Iterative improvement over multiple attempts Learning from errors without fine tuning Complex coding or decision making tasks When single pass reasoning is insufficient How It Works After task failure, the agent generates a verbal "reflection" analyzing what went wrong. This reflection is stored in memory and used in subsequent attempts to avoid repeating mistakes. Prompt Template Initial Attempt : Reflection : Subsequent Attempt (with memory) : Example: Code Generation Strengths Learns from errors without weight updates Achieves 91% on HumanEval (surpassing GPT 4's 80%) Builds episodic memory of insights Limitations Requires multiple attempts Memory management for long sessions Quality of reflection affects improvement Decision Matrix: Which Technique to Use Best Practices 1. Start Simple Begin with Zero shot CoT ("Let's think step by step"), then progress to more complex techniques if needed. 2. Match Technique to Task Math/Logic : CoT, PAL, Self Consistency Multi hop QA : ReAct, Least to Most Creative/Puzzles : Tree of Thoughts Iterative Tasks : Reflexion 3. Combine Techniques Techniques are often complementary: ReAct + Self Consistency for robust factual answers ToT + PAL for complex computational exploration Least to Most + Reflexion for hard multi step problems 4. Prompt Engineering Tips Use clear step markers ("Step 1:", "First,", etc.) Include diverse exemplars covering edge cases Format consistently across examples Add verification steps ("Let me verify...") Common Mistakes Mistake Why It's Wrong Fix Using CoT for simple lookups Adds unnecessary tokens and latency Reserve for multi step reasoning Too few samples in Self Consistency Majority voting needs adequate samples Use 5 10 samples minimum Generic "think step by step" without checking output Model may produce irrelevant reasoning Validate reasoning quality, not just presence Mixing techniques without understanding trade offs Computational cost without benefit Understand when each technique adds value Using PAL without code interpreter Code generation is useless without execution Ensure execution environment available Not testing exemplar quality in few shot CoT Poor exemplars lead to poor reasoning Validate exemplars solve problems correctly Applying Tree of Thoughts to linear problems Massive overhead for no benefit Use ToT only when exploration needed References 1. Wei, J. et al. (2022). "Chain of Thought Prompting Elicits Reasoning in Large Language Models." [arXiv:2201.11903](https://arxiv.org/abs/2201.11903) 2. Kojima, T. et al. (2022). "Large Language Models are Zero Shot Reasoners." [arXiv:2205.11916](https://arxiv.org/abs/2205.11916) 3. Wang, X. et al. (2022). "Self Consistency Improves Chain of Thought Reasoning in Language Models." [arXiv:2203.11171](https://arxiv.org/abs/2203.11171) 4. Yao, S. et al. (2023). "Tree of Thoughts: Deliberate Problem Solving with Large Language Models." [arXiv:2305.10601](https://arxiv.org/abs/2305.10601) 5. Zhou, D. et al. (2022). "Least to Most Prompting Enables Complex Reasoning in Large Language Models." [arXiv:2205.10625](https://arxiv.org/abs/2205.10625) 6. Yao, S. et al. (2022). "ReAct: Synergizing Reasoning and Acting in Language Models." [arXiv:2210.03629](https://arxiv.org/abs/2210.03629) 7. Gao, L. et al. (2022). "PAL: Program aided Language Models." [arXiv:2211.10435](https://arxiv.org/abs/2211.10435) 8. Zhang, Z. et al. (

thought-based-reasoning