Few-Shot Learning
AI technique providing a small number of task examples in the prompt to guide model performance.
Few-shot learning in the context of LLMs refers to providing a small number of labeled input-output examples (typically 2–10) directly within the prompt, allowing the model to learn the task format, expected output style, and level of detail required from these examples before performing the actual task. Unlike fine-tuning (which updates model weights), few-shot learning uses examples purely as in-context guidance at inference time.
Few-shot prompting structure: [Example 1 Input] → [Example 1 Output] | [Example 2 Input] → [Example 2 Output] | [Actual Task Input] → ?
The model infers the pattern from examples and applies it to the new input. Carefully selected examples improve performance by demonstrating: output format (JSON structure, table format, summarization length), the level of detail expected (what to include and exclude), handling of edge cases, domain-specific vocabulary usage, and uncertainty expression (when to say 'not specified' vs. providing an estimate).
For financial document processing, few-shot examples are invaluable: showing the model exactly what revenue figure to extract from an earnings release (including where to look and how to format it), demonstrating how to classify contract clauses (with examples of each category), or teaching the model the expected precision and rounding for financial metrics.
Few-shot example quality and selection matters significantly. Examples should: represent the distribution of actual task inputs (including edge cases), demonstrate the desired output format precisely, avoid ambiguity that could mislead the model, and be recent enough that they don't bias the model toward outdated patterns. Example diversity (showing the model different valid input variations) outperforms examples that are all similar to each other.
FAQs
How many examples are needed for effective few-shot prompting?
The optimal number of examples depends on task complexity, model capability, and available context window. Simple, well-defined tasks (straightforward classification, standard format extraction) often work well with 2–3 examples. More complex tasks (multi-step reasoning, nuanced classification, format with many fields) benefit from 5–10 examples. Very long documents reduce the practical number of examples because each example consumes significant context window. Research shows diminishing returns beyond 8–10 examples in most cases—more examples rarely improve performance proportionally and may introduce conflicting patterns. The quality and representativeness of examples matters far more than the exact number; a few diverse, precisely specified examples outperform many similar or ambiguous ones.
What makes a good few-shot example for financial document extraction?
Effective few-shot examples for financial extraction have: exact input text (the same document structure the model will encounter), precisely formatted output (exactly the JSON structure, field names, value formats expected), handling of ambiguous cases (demonstrating what to do when a field isn't present or is unclear—null vs. 'N/A' vs. excluding the field), numerical formatting specifications (number of decimal places, handling of thousands separators, currency symbols), and representation of edge cases likely to appear in production (non-standard date formats, negative values, scaled values where '2.5B' means 2,500,000,000). Examples that don't represent real edge cases provide false confidence and fail at deployment time when those cases appear.
Can few-shot learning work for classifying financial risk factors?
Yes—few-shot learning is effective for financial risk factor classification, particularly when categories are well-defined and examples clearly distinguish between them. For a model classifying 10-K risk factor paragraphs by type (market risk, credit risk, operational risk, regulatory risk, etc.), providing 2–3 labeled examples per category with clear, representative text demonstrating each category's defining characteristics improves classification accuracy substantially. The key challenge is: categories may have fuzzy boundaries (a paragraph can involve both market and credit risk), real examples may not be cleanly representative, and the model's training data may not align well with the company's specific taxonomy. Human-reviewed 'gold standard' examples that clearly represent each category are worth the investment for production systems.
Related Terms
Zero-Shot Learning
AI capability to perform tasks on categories or domains not seen during training, using semantic knowledge.
Prompt Engineering
Craft of designing and optimizing inputs to AI language models to reliably produce desired outputs.
Fine-Tuning
Further training a pre-trained AI model on domain-specific data to improve performance on specialized tasks.
Large Language Model
AI system trained on vast text data to understand and generate human language across many tasks.
