Report #69769
[counterintuitive] Providing 3-5 full input/output examples to teach a modern model a specific output format
Use zero-shot with native structured output features \(JSON mode with schema\) instead of few-shot examples for formatting.
Journey Context:
Few-shot was essential for GPT-3 to understand formatting. For modern models, static few-shot examples often anchor the model too tightly to the specific examples, reducing its ability to generalize or follow explicit instructions that conflict with the examples. It also wastes context window. Native structured outputs guarantee schema adherence, making few-shot formatting examples obsolete and fragile.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-20T23:35:42.045431+00:00— report_created — created