Skip to content

unexpected XML syntax from LLM causes agent to output ">\</final-answer>" after first step #152

@Music4Dogs

Description

@Music4Dogs

Issue Type

Agent Issue

Operating System

Windows

Description of the bug

When asking multiple different models up to and around 24b, they will often interpret the system prompt as though the "<final-answer></final-answer>" syntax should be output exactly as shown here, with no newline. This appears to break the agent, causing it to output "></final-answer>" as a final response and stop the agent's workflow.

Steps to Reproduce

  1. Load a small language model. (InternVL-3.5-14b, 8b, and magistral small 24b were tested)
  2. make a simple request that should result in a multi-step agent workflow, such as navigating to two separate pages sequentially.

Screenshots / Videos

Image Image

BrowserOS Version

137.0.7215.69

Additional Context

After updating to latest, none of the models i am capable of running locally will perform a multi-step workflow due to their interpretations of the system prompt not aligning with the expected response format, and presumably breaking the expected xml format. This could be solved by including more detailed clarification of the expected response structure in the system prompt, or adding the option to manually edit the agent's prompt to allow users to account for quirks with language models the BrowserOS team can't individually rewriting the system prompt for.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions