Advanced Prompting Techniques

Chain-of-Thought Prompting

CoT leverages the transformer's sequential processing. By forcing the model to generate intermediate reasoning steps, you increase the effective "thinking time" (number of forward passes through the attention mechanism). Research from Google shows CoT improves accuracy by 30-50% on arithmetic and logical reasoning tasks.

CoT Implementation Patterns

• Zero-shot CoT: Add "Think step-by-step" or "Show your work." Effective for straightforward multi-step problems. No examples needed.
• Structured CoT: Provide explicit steps: "1) Extract variables 2) Apply formula 3) Verify units 4) State answer." Use for domain-specific reasoning (financial analysis, scientific calculations).
• Self-consistency: Generate 3-5 CoT paths, select most common answer. Reduces errors from single reasoning path. Used in production for high-stakes decisions.
• Verification step: End with "Review your reasoning for errors." Model catches 20-30% of its own mistakes.

Few-Shot Learning

Few-shot learning provides task-specific training data in the prompt context. Models adjust their distribution based on examples without parameter updates. This is in-context learning: the model's behavior changes for that session only. Research shows 3-5 high-quality examples often match fine-tuned model performance for specific tasks.

Including reasoning increases accuracy by 15-20% (internal testing). The model learns not just the what, but the why. This is few-shot CoT: combining both techniques.

Few-Shot Engineering

• Example selection: Choose edge cases over obvious ones. Model already handles "Invoice question -> Billing." Show ambiguous cases that define boundaries.
• Example ordering: Most recent example has strongest influence. Place most similar example last.
• Format consistency: Exact delimiter consistency matters. "Input:" vs "Input -" changes pattern matching. Use same structure across all examples.
• Optimal count: 3-5 examples for most tasks. More examples improve accuracy with diminishing returns. Context window limits apply (examples use tokens).
• Dynamic few-shot: For production, retrieve similar examples from database based on input. This is retrieval-augmented prompting.

Technique Composition

CoT and few-shot are composable. Few-shot CoT (providing examples that include reasoning) outperforms either technique alone. Research paper "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models" demonstrates this empirically.

Other advanced combinations: Self-consistency + Few-shot (generate multiple few-shot reasoning paths), Structured CoT + Examples (provide reasoning template + filled examples). Choose based on task complexity and accuracy requirements.