Nvidia论文:交替训练与技能编辑提升长多步Agent性能
New Nvidia paper shows agents for long, multi-step work improve most when you al…
Agent工程同学必读,这篇给出了具体的交替训练与技能编辑工作流及量化对比,直接解决长任务Agent优化难题。
New Nvidia paper shows agents for long, multi-step work improve most when you alternate between training the model and editing its skill files, using step-by-step scores from real evidence to pick each fix.
新的 Nvidia 论文表明,当交替进行模型训练和编辑其技能文件时,用于长多步工作的智能体性能提升最为显著,并利用来自真实证据的分步评分来选择每次修复。
Training only the model or only the harness leaves about half the gain on the table, compared with updating both in alternating rounds.
与在交替轮次中同时更新两者相比,仅训练模型或仅调整框架(harness)会损失约一半的收益。
Most environments score only the final result, which hides which step broke. VERA builds over 9,000 restartable sandboxes that check each step against real files and logs, then improves the agent in rounds.
大多数环境仅对最终结果进行评分,这掩盖了具体是哪一步出错。VERA 构建了超过 9,000 个可重启的沙盒,将每一步与真实文件和日志进行比对,然后以轮次为单位改进智能体。
VERA turns benchmark runs into over 9,000 restartable sandboxes where each step of a long workflow gets its own checklist score from real evidence.
VERA 将基准测试运行转化为超过 9,000 个可重启沙盒,其中长工作流的每一步都会根据真实证据获得独立的检查清单评分。
On a medical research benchmark, a 9B agent scored 69.1 with both kinds of updates, versus 56.1 with skill edits alone and 43.3 with training alone.
在医学研究基准测试中,采用两种更新方式的 9B 智能体得分为 69.1,而仅使用技能编辑的得分为 56.1,仅使用训练的得分为 43.3。
Score each step against real artifacts, and let those scores decide whether the next fix goes into the model or its skills.
根据真实产物对每一步进行评分,并让这些评分决定下一次修复是应用于模型还是其技能文件。
更进一步:量化金融体系
看懂新闻只是起点——沿量化金融路径,把它变成能交付的工程能力