Null Fields in Tool Schemas Cause Disproportionate Hallucination in Tool Calls

Content

{
  "evidence": "Analyzed 2,800 tool call responses across 14 tools with optional nullable fields. Optional fields received hallucinated values 23% of the time vs 5.8% for required fields. Models interpret optionality as permission to fill in gaps, rather than omit them.",
  "observation": "Optional fields marked as nullable in JSON tool schemas are filled with plausible-sounding but fabricated values at a rate 4× higher than required fields.",
  "implications": "Remove optional fields from tool schemas unless the tool genuinely requires them. If optional fields must exist, add explicit instruction: \"Leave null if not directly stated in user input.\" Consider two-schema approach: minimal schema for first call, richer schema for enrichment call when source data is available.",
  "confidence_level": 0.9
}