Success Is Not Self-Explanatory: Auditing Success Provenance in Agent Evaluation
2026-07-27 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors point out that just getting the right answer doesn't tell us how an AI agent arrived at it or if it used the intended information correctly. They created AcquaBench, a method that compares different versions of test inputs to see if success really depends on the right data or just exposure to source structure. Their findings show that sometimes the agent’s success is linked to using the correct information, but other times behavior can persist even when the direct clue is removed. They suggest benchmarks should not only report if agents succeed but also if the information they used actually supports that success.
agent evaluationsuccess provenanceAcquaBenchbenchmark testinginformation statevalue substitutionCLEAN conditionGOLD conditionSHAM conditionAUROC
Authors
Jingkun Luo, Da-Tian Peng
Abstract
A correct answer can conceal why an agent succeeded. Once agents change their information state during evaluation, correctness no longer distinguishes intended reasoning from answer acquisition. Outcome evidence and exposure detection do not establish whether success depended on an acquired target; we call this missing evaluation object success provenance. AcquaBench audits it through matched CLEAN, GOLD, and SHAM value substitution on four standardized surfaces with joint qid-clustered analysis. CLEAN retains benchmark-authorized information. GOLD makes the correct target available. SHAM preserves source structure and exposure opportunity but substitutes a matched incorrect value. GOLD minus CLEAN measures the total score response to correct-target availability; GOLD minus SHAM tests whether that response tracks target correctness beyond matched source exposure. In D0, GOLD exceeds SHAM by 19.1 to 25.9 percentage points, showing that success follows the correct value. In D2, GOLD still exceeds SHAM under distributed sufficiency while coloc no longer transfers as a high-score marker, with AUROC 0.376 and 0.142. Behavioral dependence can thus persist beyond this probe's intended observation unit. In model comparison, a supported 5.0-point CLEAN score gap compresses to a raw GOLD difference of -0.6 points without establishing rank inversion. Agent benchmarks should report success together with whether the evaluated information state supported it.