Supplying worked examples in the prompt.
Few-shot has not stopped working. It has stopped working for reasoning on models that reason natively — which is a much narrower claim than the one circulating, and the difference matters when you decide what to delete.
What survives is format alignment. If you need output shaped a particular way, examples remain the cheapest instrument available. If you are paying for examples hoping they improve the model’s thinking, check the ledger below for your model family first.
few-shot · domain-multimodal · gpt-4o
recommended confidence: moderate no check recorded
52.17% zero-shot to 73.52% at 16-shot
Can Multimodal LLMs See Materials Clearly? A Multimodal Benchmark on Materials Characterization · arXiv preprint, 2025 not read in full
Contested by: few-shot-cot--reasoning--qwen2-5
few-shot · output-formatting · all
recommended confidence: moderate no check recorded
Still the cheapest way to pin an output schema
Revisiting Chain-of-Thought Prompting: Zero-shot Can Be Stronger than Few-shot · EMNLP 2025 Findings, 2025 not read in full