How hidden scaffolding and flawed scoring distort AI agent benchmarks

Agent benchmarks are widely relied upon to evaluate and compare large language models, yet a fundamental flaw undermines their validity: these scores often measure the evaluation pipeline itself rather than genuine model capability.

Researchers have identified what they call a "double measurement confound" that distorts benchmark results in two ways. First, scaffolding—support structures built into the test environment—silently handles critical execution decisions that should be performed by the model being evaluated. Second, scoring mechanisms evaluate model outputs using surface-level criteria rather than checking against ground truth. Together, these issues can make weak models look competitive or hide real performance gaps.

The hidden scaffolding problem#

When benchmarks test agent behavior, they typically include helper systems and support layers that guide or enable execution. The problem arises when the scaffold—not the model—actually makes key decisions. This shifts what's being measured: instead of assessing the model's reasoning and execution ability, the benchmark ends up rating the quality of the support structure. Different models tested with the same benchmark may benefit differently from the scaffolding, making direct comparison misleading.

Scoring that ignores ground truth#

Evaluation typically relies on shape-based scoring, where outputs are assessed against surface patterns or predetermined rubrics. This approach misses when an answer is genuinely correct or incorrect. Seeded ground-truth scoring—checking outputs against known-correct solutions—provides a more accurate picture of what models can actually achieve.

A framework for better benchmarks#

The researchers propose an audit-and-repair protocol that addresses both problems. The approach transfers critical decisions from scaffold to model, replaces pattern-matching scoring with ground-truth validation, and reports reliability metrics beyond average performance—including worst-case scenarios and tail-risk measures. Testing this approach on ComtradeBench revealed that what appeared as a nearly flat leaderboard became a meaningful spectrum showing real differences in both average capability and robustness across test variations.

The work suggests that benchmark designers should disclose scaffolding levels, scoring criteria, and reliability profiles alongside headline scores, enabling more informed interpretation of model capability claims.