Bug Report: Tool Call XML Label Hallucination in Multi-turn Reasoning (Nemotron-3 Super 120B)
Description
There is a consistent XML syntax error during tool-assisted reasoning. The model intermittently generates a "merged" or "hallucinated" tag <function=parameter=code> instead of following the standard nested structure defined in the ChatML template.
Patterns Observed
Instead of the canonical format:
<tool_call>
<function=python>
<parameter=code>
{python_code}
</parameter>
</function>
</tool_call>
The model frequently outputs:
<tool_call>
<function=parameter=code>
{python_code}
</parameter>
</function>
</tool_call>
Note the malformed opening tag <function=parameter=code>.
Impact
This behavior breaks standard parsers (including vLLM's Qwen3CoderToolParser). In complex tasks like mathematical reasoning, the model seems to lose track of the XML hierarchy after a long thought process, leading to execution failures in automated pipelines.
Environment & Context
- Model:
NVIDIA-Nemotron-3-Super-120B-A12B-FP8 - Prompting Style: ChatML with tool-calling enabled.
- Task: Multi-step mathematical reasoning involving Python tool calls.
I ran into a similar issue with it today, running fairly simple calls.
<function=Get-ChildItem>
<parameter=Path>
C:\Users\xxxx\xxxx
<parameter=Force>
True
Had to switch off of it, as all other models perform these calls fine.
Facing the same issue with vLLM Deployment. Did anyone found any fix to use tool calling with this model ?