Back to all discussions

  •   11 months ago

Gemini 2.5 Pro tool calling issue

I am wondering if someone from Google is actively looking here? Gemini 2.5 Pro has a recurring issue: after initially working fine as an ADK model, it suddenly starts hallucinating tool argument names in a very specific way. For example:
user_input becomes usera_input
theme_and_message becomes themea_anda_message
test_arg becoms testa_arg

At this point I switch to gpt-4o to continue working, but it is truly frustrating.

  • 6 comments

  •   •   11 months ago

    I guess nobody is watching here but this is a truly frustrating issue will ALL Gemini models:

    Another example:
    def edit_campaign_data_form(campaign_data: dict) -> dict:

    The tool expects the argument to be named campaign_data, but Gemini keeps calling it campaigna_data — causing this error:
    Invoking edit_campaign_data_form() failed as the following mandatory input parameters are not present: campaign_data

  • Manager   •   11 months ago

    Hi Jeny,

    The Google Team is not checking these, but the Devpost team will report technical issues to them. You can also report errors or typos in the adk-docs tutorial is to file a GitHub issue. This ensures the team responsible for the documentation sees your feedback directly.

    The direct link for filing a GitHub issue for the google/adk-docs repository is:

    https://github.com/google/adk-docs/issues/new

    Even better, if you're comfortable with it, you could submit a pull request with your suggested corrections to github.com/google/adk-docs. This is often the quickest way to get fixes implemented, as it allows you to contribute the solution directly.

  •   •   11 months ago

    Hi Shawni,

    This issue is likely not related to Google ADK but to how the LLM (Gemini) handles tool calling and is definitely something I cannot fix.
    I doubt that I am alone in experiencing this issue, but it is something which cannot be easily reproduced. I will see where to submit it.

    Thanks!

  • Manager   •   11 months ago

    Ah okay thanks for the clarification. I've raised it to the Google Team and will let you know when I hear back from them.

  • Manager   •   11 months ago

    Hi again. I've just received the following response from the Google Team: "Thank you for sharing your experience. I completely understand how frustrating it must be to consistently encounter this specific tool argument hallucination with Gemini 2.5 Pro in ADK, especially after it initially works well.

    I strongly recommend opening a detailed issue on the official Google ADK GitHub repository (https://github.com/google/adk-python/issues) or the Vertex AI issue tracker (https://issuetracker.google.com/issues/new?component=1130925&template=1637248). This ensures it gets directly to the product and engineering teams responsible for the models and ADK. Includ e a minimal reproducible example: The most effective way for our engineering team to investigate and address this is with a minimal, self-contained code example that reliably reproduces the issue. This would include:
    - Your ADK agent definition (especially the LlmAgent and how tools are registered).
    - The exact FunctionTool definition that is failing.
    - The input prompt you are providing that triggers the hallucination.
    - The full error message you receive.
    - The google-cloud-aiplatform and ADK Python SDK versions you are using."

  •   •   11 months ago

    Thank you Shawni, I will report this.

Comments are closed.