complex cot

**Complex CoT (Complex Chain-of-Thought)** refers to chain-of-thought prompting techniques specifically designed for **multi-step, difficult reasoning problems** — using longer, more detailed reasoning chains, richer demonstration examples, and structured decomposition to handle problems that simple CoT fails to solve. **Why "Complex" CoT?** - Standard CoT with short reasoning traces works well for simple problems (basic arithmetic, single-step logic). - **Complex problems** — involving many reasoning steps, multiple sub-problems, or requiring integration of different knowledge types — need **more elaborate reasoning chains** to succeed. - Complex CoT provides these longer, more structured chains either through carefully designed prompts or through techniques that encourage deeper reasoning. **Complex CoT Techniques** - **Longer Demonstrations**: Use few-shot examples with **detailed, multi-step reasoning** — 10–20 reasoning steps per example rather than 3–5. - **Complexity-Based Selection**: When choosing few-shot examples, **prioritize complex examples** over simple ones — research shows that demonstrations with more reasoning steps produce better results even on simpler test questions. - **Multi-Path Reasoning**: Generate multiple reasoning paths and combine them: - **Self-Consistency**: Sample many CoT traces, take majority vote on the answer. - **Multi-Chain**: Different prompts or decomposition strategies, ensemble the results. - **Hierarchical Reasoning**: Break the problem into sub-problems, solve each with its own CoT, then combine: ``` Main Problem: [complex question] Sub-problem 1: [simpler aspect] CoT for sub-problem 1: ... Sub-answer 1: ... Sub-problem 2: [another aspect] CoT for sub-problem 2: ... Sub-answer 2: ... Final reasoning: Combining sub-answers... Final answer: ... ``` **Complex CoT for Different Domains** - **Mathematics**: Multi-step proofs and derivations — each step building on the previous, with explicit justification. - **Programming**: Algorithm design → pseudocode → implementation → testing → debugging — structured development chain. - **Scientific Reasoning**: Hypothesis → evidence evaluation → mechanism analysis → conclusion — scientific method as CoT. - **Legal/Policy Analysis**: Rule identification → fact mapping → precedent analysis → conclusion — structured legal reasoning. **Complexity-Based Prompting (Key Finding)** - A key research finding: selecting few-shot examples based on **reasoning complexity** (number of steps in the solution) outperforms selecting examples based on similarity to the test question. - Using the **most complex available examples** as demonstrations encourages the model to reason more thoroughly — even when the test question is simpler. - This suggests that complex demonstrations teach the model **how to reason deeply** rather than just providing task-specific patterns. **Benefits of Complex CoT** - **Harder Problems**: Handles problems that simple CoT cannot — multi-hop reasoning, multi-constraint satisfaction, complex calculations. - **Better Calibration**: Longer reasoning chains give the model more opportunity to catch and correct errors. - **Richer Explanations**: The detailed reasoning provides more interpretable and verifiable traces. Complex CoT represents the **frontier of prompted reasoning** — it pushes the boundaries of what language models can solve through carefully structured, multi-step reasoning chains.