[Bug] Structured outputs float generation leads to infinite zeros

[inogii]

Hi,

I have been trying to apply qwen3-coder to a mathematical reasoning use case, since as far as my understanding goes, it should be better in this task than a standard qwen model.

As a first step I am trying to make the model extract variables with their values from a given prompt. I am using a structured outputs approach, where i give a pydantic model's json schema with the structure of the variable. One of the fields is a float value, representing the value of the variable. You can see the schema below:

class Variable(BaseModel):
    symbol: str = Field(
        description="Short algebraic symbol used inside equations, e.g. 'v', 'm', 'KE'."
    )
    description: str = Field(
        description="Human-readable description of what this variable represents."
    )
    unit: Optional[str] = Field(
        default=None,
        description="Physical or financial unit, e.g. 'm/s', 'kg', 'USD'. Omit if dimensionless."
    )
    known_value: Optional[float] = Field(
        default=None,
        description=(
            "Numeric value if the problem statement provides it. Do not add trailing zeros - e.g. write 0.5 instead of 0.500"
            "Set to null if this variable is unknown and must be derived."
        )
    )

Whenever I try this approach with the Qwen3-Coder-Next model, it is outputting an infinite list of zeros for the known_value field. I have tried to solve it with prompting, but I have not been successful. I find it very curious, since when testing other models of the Qwen3 family (more specifically Qwen3-VL-30B-A3B-Instruct) it has worked flawlessly, and this coding model that should be better at mathematics related tasks is hallucinating a lot. You can see a snippet of the output I get below:

  Example [0]: Kinematics – final velocity

  Problem:
    A car starts from rest and accelerates uniformly at 3 m/s² for 8
    seconds. What is the final velocity and how far does it
    travel?

  Streaming model output (raw JSON tokens):
  ────────────────────────────────────────────────────────────────────

{
  "problem_summary": "A car starts from rest and accelerates uniformly at 3 m/s² for 8 seconds. Find final velocity and distance traveled.",
  "variables": [
    {
      "symbol": "u",
      "description": "Initial velocity",
      "known_value": 0.0000000000000000e+000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000000

Is there any known solution to this kind of problem? Thanks in advance.