Prompting Techniques by Task Type

A survey of 39 prompting techniques across 29 NLP tasks reveals distinct patterns in what works where. No single technique dominates; optimal selection depends on task characteristics.

Mathematical Problem Solving

Best approaches: Program-of-Thoughts (PoT), Program-Aided Language Models (PAL), Chain-of-Code

Why: These tasks require precise computation where LLMs make arithmetic errors. Delegating calculation to external interpreters eliminates this failure mode while leveraging the model’s strength at problem decomposition.

Performance: PoT shows ~12% average improvement over Chain-of-Thought; PAL achieves >90% on some reasoning benchmarks.

Reasoning Tasks (Logical, Commonsense, Multi-hop)

Best approaches: Chain-of-Thought, Self-Consistency, Tree-of-Thoughts

Why: Reasoning benefits from explicit intermediate steps. Self-Consistency adds robustness by sampling multiple paths. Tree-of-Thoughts enables exploration with backtracking for complex problems.

Performance: CoT improves mathematical reasoning by up to 39%, commonsense by 26%. Tree-of-Thoughts shows 65% better success rates than CoT on mathematical tasks.

Knowledge-Intensive Tasks (QA, Fact Verification)

Best approaches: Chain-of-Verification, Implicit RAG, Chain-of-Knowledge

Why: Factual accuracy requires verification mechanisms. These methods either check claims against external knowledge or have the model verify its own assertions through independent questioning.

Performance: CoVe improves F1 by up to 23% on closed-book QA.

Classification and NLU Tasks (NER, Relation Extraction, Inference)

Best approaches: Metacognitive Prompting

Why: These tasks benefit from nuanced interpretation and self-reflection rather than step-by-step reasoning or computation. The five-stage metacognitive process (comprehend → judge → evaluate → decide → assess confidence) consistently outperforms alternatives.

Performance: Up to 26.9% improvement on domain-specific classification tasks.

Table-Based Reasoning

Best approaches: Chain-of-Table, specialized structured prompts

Why: Tabular data has inherent structure that general reasoning methods don’t exploit effectively. Techniques designed for table manipulation consistently outperform general methods.

When Each Applies

Task Type	Primary Method	Why It Works
Math/numerical	PoT, PAL	Offloads error-prone computation
Multi-step reasoning	CoT, ToT	Surfaces intermediate steps
Factual accuracy	CoVe	Self-verification catches errors
Classification/NLU	Metacognitive	Self-reflection improves judgment
Structured data	Domain-specific	Exploits inherent structure

The pattern: match the prompting technique to the primary challenge of the task, computation, reasoning, verification, interpretation, or structure.

Related: 05-molecule—chain-of-thought-prompting, 05-molecule—tool-delegation-pattern, 05-molecule—metacognitive-prompting, 05-molecule—chain-of-verification