LLM agent evaluations can miss errors in tool usage and verification
What the Trace Shows and What the Label Says: Auditing Behavioral Detectors in Agent Evaluation
Artificial IntelligenceSoftware Engineering
Summary
Evaluations of AI agents often rely on labels that summarize their behavior, but these labels might miss important details. The authors show that AI models sometimes silently correct errors from tools they use without showing it in verification checks. By comparing automated labels with human reviews, they find that some checks are accurate while others miss common patterns. They suggest showing both the labels and the detailed trace evidence to better understand the agent’s behavior during evaluation.
What this means in practice
- •For ai system evaluators: Improve audit processes by accessing trace evidence alongside automated labels to better detect model errors and tool output misuse.
- •For software developers using ai tools: Use detailed trace inspection tools to understand and debug agent interactions with external tools during development.
Authors
Idunnuoluwa A. Adeniji, Shilpika, Michael E. Papka
Abstract
LLM agents are often evaluated by reducing execution traces to labels such as final-answer correctness, tool verification, and fault propagation. These labels are measurements produced by parsers, behavioral detectors, and harness rules, and they may not fully capture what the trace shows. We study this in a tool-use setting built on GSM8K, where agents call a calculator whose outputs are occasionally corrupted, giving clear ground truth for what the tool returned, what the model later used, and what the detectors reported. We audit three trace-level detectors across six production models and 30 problems per model. Final correctness, visible re-verification, and fault propagation can disagree in systematic ways. A key failure mode is silent tool-output override: the calculator returns a corrupted value, but the model later uses a different, often correct, value without calling the tool again, so the bad value stops propagating while a verification detector records no re-check. Human review of 90 traces agrees perfectly with the automated labels for final correctness (67/67), fault impact (19/19), and fault relevance (18/18), but only 67/81 for verification; all 14 disagreements fall into four recurring patterns. Execution limits also change the reported result: in an ablation across two models, one ranges from 21/30 to 27/30 correct as step and token budgets change, while the other stays between 27/30 and 29/30. These findings motivate measurement transparency: an automated label should come with the trace evidence behind it, whether the question the detector asks applied to that trace, whether the run was cut off by an execution limit, and any human corrections. Our trace viewer makes this information inspectable alongside the reported results.