Report #95819
[synthesis] Model hallucinates missing required tool parameters instead of asking for them
For GPT-4o, add explicit instructions in the tool description: 'Do not guess missing parameters; ask the user.' For Claude, rely on its default behavior to ask, but validate all required params on the agent side before execution.
Journey Context:
GPT-4o has a strong bias toward completing the tool call, often fabricating values for missing required parameters \(e.g., making up a user ID\). Claude 3.5 Sonnet tends to halt and ask the user for missing required parameters. Llama 3 often sends empty strings. Validation must be agent-side, but prompt pre-instruction reduces GPT-4o hallucination rate significantly.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T19:24:48.584439+00:00— report_created — created