Why Do LLM Function Calls Hallucinate Parameters and How to Prevent It?

1. Why Function Calls Hallucinate Parameters

The model’s default behavior follows a maximum‑likelihood bias: when information is missing it tries to fill the gap, generating plausible but incorrect values. Because Function Calls are still generated text, the model may invent parameters just as it would produce a fluent sentence.

2. Five Typical Hallucination Patterns

Self‑created parameters (e.g., fabricating flight_id or hotel_id).

Implicit default values (e.g., inserting "tomorrow" when no date is provided).

Cross‑turn contamination (carrying a date from a previous turn into a new, unrelated request).

Wrong tool selection (answering a weather query with a flight search).

Incorrect tool call order (booking a ticket before searching for flights).

If uncontrolled, these issues render the system unusable.

3. Systematic Ways to Reduce Hallucination (Engineering, Not Prompt Tricks)

Schema design.

Prompt context.

Dynamic tool routing.

Result‑validation layer.

Clarification (follow‑up questioning) mechanism.

The following sections detail each direction using a travel‑assistant case study.

4. Method 1 – Unambiguous, Non‑Default Schema

A robust schema must be "no ambiguity, no defaults, no vagueness." For example, a naïve schema that declares fields as simple strings leaves the model unsure whether "tomorrow" or "next Friday" counts as a valid value, leading to hallucination.

{
  "origin": "string",
  "destination": "string",
  "date": "string"
}

Improved schema adds explicit type and description, forcing the model to ask when a required value is missing:

{
  "origin": {
    "type": "string",
    "description": "User‑explicit city; do not guess. If missing, ask the user."
  },
  "destination": {
    "type": "string",
    "description": "User‑explicit city; do not guess."
  },
  "date": {
    "type": "string",
    "description": "Exact travel date; if absent, request clarification."
  }
}

5. Method 2 – Prompt Rule: No Guessing, Must Ask

The standard prompt includes a clear instruction:

Do not guess missing information; always ask the user for clarification.

Correct behavior example: Assistant: "Which city are you departing from?" Incorrect behavior (hallucination) example:

search_flights(origin="Shanghai", destination="Beijing", date="tomorrow")

Adding this rule typically cuts error rates by 20‑40%.

6. Method 3 – Dynamic Function Routing

When many tools are available, the model may pick the wrong one. Dynamic routing first filters the tool set based on the user query, then only presents the remaining candidates to the model.

Example: user asks "What's the weather in Beijing tomorrow?" – routing reduces the options to get_weather only, preventing accidental flight searches. Real‑world tests show a single‑tool routing can reduce wrong calls by up to 50%.

7. Method 4 – Result Validation Layer

After a tool call, a backend validator checks three aspects:

Parameter completeness (e.g., missing origin).

Schema conformity (type and description constraints).

API‑level errors (HTTP failures, malformed responses).

When a missing parameter is detected, the validator returns an error such as: Error: missing origin The model then asks the user to provide the missing value, turning a fatal hallucination into a recoverable interaction.

8. Method 5 – Clarification Mechanism

In practice, missing parameters are the norm. The system forces the model to ask for each required field before invoking any tool:

Assistant: "Please tell me the departure city, destination, and travel date."

In the training‑camp projects, this clarification step reduced parameter hallucination by nearly 70%.

9. Real‑World Case Study – Five‑Step Optimization

Initial model output (hallucinating "Shanghai"):

search_flights(origin="Shanghai", destination="Beijing", date="tomorrow")

Five corrective steps applied:

Schema tightening : add "origin cannot be guessed" description.

Prompt rule : enforce "ask when missing".

Dynamic routing : limit tools to relevant ones.

Validation layer : reject calls missing required fields.

Clarification : ask the user for the origin.

After these steps the assistant correctly asks:

To book your flight, could you confirm the departure city?

The system becomes stable and production‑ready.

10. Interview Answer Outline

When asked "How to avoid Function Call hallucination?" respond with the five engineering pillars above: clear schema, non‑guessing prompt, dynamic routing, backend validation, and a clarification mechanism.