hermes - 💡(How to fix) Fix [Resilience] REASONING_SCRATCHPAD false positive detection triggers retry loop [2 comments, 2 participants]

1. Context contamination: When the conversation context contains code examples, error logs, or tool outputs that include the literal string <REASONING_SCRATCHPAD> (e.g., from searching source code), the model may reference this text in its response, triggering the detector.

TL;DR

Implementing a more robust check for incomplete scratchpad, such as the suggested fix, can help mitigate false positives by properly handling context contamination and ensuring accurate XML structure validation.

Guidance

• Modify the has_incomplete_scratchpad function to strip code blocks and validate XML structure, as suggested, to reduce false positives.
• Verify the effectiveness of the fix by testing scenarios that previously triggered false positives, such as conversations with code examples or error logs containing the literal string <REASONING_SCRATCHPAD>.
• Consider adding additional logging or monitoring to track the frequency of false positives before and after implementing the fix to measure its impact.
• Review the verification_gate_retries mechanism to ensure it is appropriately configured to handle legitimate retries without dropping responses unnecessarily.

Example

The provided code snippet in the issue body demonstrates a potential solution:

def has_incomplete_scratchpad(text: str) -> bool:
    import re
    cleaned = re.sub(r'```.*?```', '', text, flags=re.DOTALL)
    cleaned = re.sub(r'`[^`]*`', '', cleaned)
    opens = len(re.findall(r'<REASONING_SCRATCHPAD>', cleaned))
    closes = len(re.findall(r'</REASONING_SCRATCHPAD>', cleaned))
    return opens > closes

This example strips code blocks and checks for the presence of opening and closing tags to determine if a scratchpad is incomplete.

Notes

The suggested fix assumes that the <REASONING_SCRATCHPAD> tag is always properly formatted and that the issue is primarily caused by context contamination and lack of XML structure validation. Additional testing may be necessary to ensure the fix does not introduce new issues or affect other parts of the system.

Recommendation

Apply the suggested workaround by modifying the has_incomplete_scratchpad function as described, to mitigate false positives and improve the overall reliability of the response generation system. This approach directly addresses the identified causes of the issue and provides a clear, step-by-step solution to enhance the system's performance.