做 AI 模型评测或选型的人必须看——这篇研究用严格对照实验证明，你看到的模型能力分数可能更多反映的是 scaffold 设计而非模型本身，建议重新审视自己的评估流程。
一项预注册的对照研究系统比较了三种不同 scaffold（ReAct、多智能体 Planner-Actor-Rater、Planner-then-Executor）在五个模型（Claude Opus 4.7、Sonnet 4.6、Haiku 4.5、Gemini 3.1 Pro Preview、GPT-5.5）上的 GAIA 验证集表现。研究发现，仅 scaffold 选择就使同一模型的准确率波动高达 28 个百分点，证实了 scaffold 变化至少产生 10 个百分点的差距。更令人意外的是，更强大的模型并未对 scaffold 更不敏感——在更难的 Level 2 任务中，最强模型反而从结构化 scaffold 中获益最多。多智能体设计在 Anthropic 模型家族中优于 ReAct，但跨模型提供商时优势消失。结构化 scaffold 调用工具次数更少，但在困难任务中从错误中恢复的能力更强。这些结果表明，单 scaffold 的能力评估数字是 scaffold 条件性的，且随着模型改进，评估差距未必会缩小。
Scaffold Effects on GAIA: A Controlled Comparison
Published agent capability scores conflate what a model can do with what its scaffold lets it do, and the magnitude of this elicitation gap is not well characterized under controlled conditions. This study executes a pre-registered controlled comparison of three scaffolds (ReAct, a Planner-Actor-Rater multi-agent design, and planner-then-executor) across five models from three providers (Claude Opus 4.7, Sonnet 4.6, Haiku 4.5; Gemini 3.1 Pro Preview; GPT-5.5) on GAIA validation Levels 1 and 2, holding tasks and conditions fixed, with three attempts per question. Scaffold choice alone moves measured accuracy by as much as 28 percentage points within a single model (Opus, Level 2, robust slice), confirming the pre-registered hypothesis that scaffold variation produces gaps of at least 10 points. The pre-registered prediction that more capable models would be less scaffold-sensitive is rejected in direction: scaffold effects vary significantly by model in every dataset slice, but the most capable Anthropic model gains the most from structured scaffolds at the harder level, and tier-scaling holds only at Level 1 under the robust slice. The multi-agent advantage over ReAct at Level 2 appears within the Anthropic family but not for the cross-provider models, making model family rather than capability tier the conditioning variable, and the predicted planner-executor advantage on file-reading tasks is falsified. Structured scaffolds make fewer tool calls yet recover more often from mid-trajectory errors at the harder level, and a single cell (Gemini with planner-then-executor) is the cheapest at both levels and the most accurate at Level 2. These results indicate that single-scaffold capability numbers are scaffold-conditional estimates and that the elicitation gap is not guaranteed to shrink as models improve.