When Tools Lie: Reliability of Mathematical Agents Under Corrupted Tool Feedback
Corrupted tool feedback drops math-agent accuracy from 100% to 72.4%; mandatory reflection restores 100%.
The paper tests whether mathematical agents notice silently wrong tool results by intercepting selected calls and replacing outputs with plausible incorrect values. Without verification, accuracy on 31 problems falls from 100% to 72.4%. Mandatory same-context reflection restores performance to 100%, while optional fresh-context or structural checks help only if the model invokes them. After explicit detection, restarting the full problem succeeds in all recovery cases.
- Silent tool corruption drops accuracy from 100% to 72.4% on 31 problems.
- Mandatory same-context reflection fully restores accuracy to 100%.
- Optional verification helps only when the model chooses to invoke it.
- Restarting after explicit detection succeeds in every tested case.
From the original article
1 min read at arxiv.orgMathematical problem solving often requires deterministic computational steps that agents delegate to tools and implicitly trust. Yet tools can fail silently, returning plausible but incorrect results. How well can agents detect and correct corrupted tool call outputs?
By Kavienan Jegatheesan · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.