Nvidia最新发布的研究指出，AI系统的软件框架对长期任务的完成能力影响更大，甚至超过了基础模型本身的作用1。通过优化内存管理并添加"监督者"组件等框架调整，Nvidia研究人员使Claude Opus 5在ARC-AGI-3基准测试中的成绩从30%提升至100%，而不使用该框架时的表现仅为30%1。Nvidia副总裁Adel El Hallak表示："Agent不仅是模型，还包括模型周围的框架、工具集、运行时和技能库"1。
这一发现得到了业界其他机构的验证。OpenAI在其模型上进行了类似研究，通过调整两个框架设置将得分提高三倍，但最终未能达到100%的水平1。Databricks首席执行官Ali Ghodsi进一步指出："使用不同框架可能使成本增加两倍，即使使用同一模型"1。为了推广这一方向，Nvidia创建了名为Agentic Variation Operators（AVO）的增强框架，并在其Nemo品牌下提供开源技术1。
Nvidia released research demonstrating that the software framework surrounding an AI model significantly outweighs the model itself in determining performance on complex tasks.1 By implementing a custom framework with optimized memory management and a supervisory component, Nvidia researchers improved Claude Opus 5's performance on the ARC-AGI-3 benchmark from 30% to 100%, showcasing the dramatic impact of architectural choices.1
According to Nvidia Vice President Adel El Hallak, "An agent is not just a model, but also includes the framework around the model, the toolset, the runtime, and the skill library."1 This perspective challenges the prevailing focus on model selection as the primary driver of AI capability. The research indicates that while choosing the right foundational model matters, the infrastructure and tools integrated with it play a more decisive role in agent effectiveness. OpenAI conducted parallel investigations on its own models, achieving a threefold score improvement through framework adjustments, though it did not reach 100% accuracy.1
The cost implications of framework optimization are substantial. Databricks CEO Ali Ghodsi noted that "using a different framework could double costs, even when using the same model,"1 underscoring that framework engineering represents both a technical and financial consideration. Nvidia developed an enhanced framework called Agentic Variation Operators (AVO), offering open-source technology under its Nemo brand.1
评论
还没有评论，欢迎留下第一条。