Step-by-step auditing improved AI fault detection—but used 3.4 times the tokens

Asking an AI auditor to inspect an agent’s work step by step improved its ability to identify the first hallucination in the AgentHallu benchmark. GPT-5 rose from 32.7% accuracy under standard prompting to 42.7% with the step-by-step procedure, while reported token use increased from 7,426 to 25,454—about 3.4 times as much. The result gives a concrete example of a better audit carrying a higher inference cost.[1][2]

AgentHallu contains 693 multi-step agent trajectories across seven frameworks and five task domains. The authors evaluate 13 models on identifying where a hallucination begins. This is a diagnosis task: a system must locate the step that caused the problem, rather than merely decide that the final answer is wrong.[1][2]

The best score depends on the procedure

Under the paper’s standard prompting setup, Gemini 2.5 Pro leads the reported overall results at 41.1%. That figure is not the maximum across all experiments: the separate step-by-step evaluation takes GPT-5 to 42.7%. Both numbers concern localisation among hallucinated instances under the benchmark’s scoring rules, not a model’s general accuracy on all agent activity.[2]

The distinction matters for anyone building an audit system. Model choice is only one part of the design; how the model is asked to inspect the record also changes the result. In GPT-5’s case, the improvement is 10 percentage points, with roughly 3.43 times the reported tokens. More thorough inspection helped, but still left substantial room for missed or misplaced diagnoses.[2]

Tool-use errors were especially difficult in the standard evaluation. Table 4 gives an 11.6% average for the five proprietary models in that category; Gemini 2.5 Pro’s own tool-use score is 14.6%. The lower figure is a group average, not Gemini’s individual result.[2]

Why finding the first error is harder

An agent’s later work can remain internally consistent after an earlier mistake. A wrong value returned by a tool, an unsupported assumption during planning or a misread source can flow into later calculations. To diagnose the failure, an auditor has to connect the final problem to the earliest relevant step without blaming a later symptom.

The benchmark organises failures across planning, retrieval, reasoning, human interaction and tool use, with finer subcategories. That structure helps distinguish which parts of an agent workflow need better checking. A system that performs well on reasoning mistakes may still struggle when the evidence is embedded in a tool call or its response.[2]

The practical lesson is to assess both diagnosis quality and the resources required to obtain it. Token use is an informative cost indicator, although actual latency and price depend on the deployment. AgentHallu offers a reproducible test setting for that comparison; it does not yet show that automated auditing can reliably replace review in every real workflow.[1][2]

Report an error or suggest a correction