NVIDIA发布VERA:验证环境与模型协同进化
Banger paper from NVIDIA.
Agent领域关键突破,提出了模型与评测环境协同进化的新范式,数据表现强劲,值得研究Agent训练流程的同学关注。
Banger paper from NVIDIA.
来自 NVIDIA 的惊艳论文。
One exciting trend I am seeing is building verifiable environments for your agents and training the harness alongside the model.
我观察到的一个令人兴奋的趋势是,为你的智能体构建可验证的环境,并与模型同步训练评估框架(harness)。
Co-evolving the harness and the model is a big part of owning the intelligence stack. And many frontier AI companies have started doing that.
协同进化评估框架和模型是掌控智能栈的关键部分。许多前沿 AI 公司已经开始这样做。
This paper shows how this works:
这篇论文展示了其工作原理:
VERA turns benchmark trajectories into more than 9,000 restartable sandboxes with rubric scoring, and keeps only environments that run and can be scored from observable evidence.
VERA 将基准测试轨迹转化为超过 9,000 个可重启的沙箱环境,并依据评分标准进行打分,仅保留那些能够运行且可通过可观测证据进行打分的场景。
It then updates both the model weights and the harness.
随后,它同时更新模型权重和评估框架。
A harness edit is kept only if it passes self-tests and adds at least 5 points on the development set.
只有当评估框架的修改通过自测并在开发集上至少提升 5 分时,才会被保留。
The system rejects a model checkpoint if its score drops by more than 20%.
如果模型检查点的分数下降超过 20%,系统会拒绝该检查点。
At 9B, the co-evolved agent beats the strongest single-axis baseline by 10.3 points on AutoCoWorkBench and 13.0 points on AutoMedBench. At 27B, it scores 71.6 on AutoCoWorkBench, above Claude Opus 4.8.
在 9B 参数规模下,协同进化的智能体在 AutoCoWorkBench 上比最强的单轴基线高出 10.3 分,在 AutoMedBench 上高出 13.0 分。在 27B 参数规模下,它在 AutoCoWorkBench 上得分为 71.6,高于 Claude Opus 4.8。
The environment corpus is open-sourced.
环境语料库已开源。
Paper: https://arxiv.org/abs/2610.05923
论文:https://arxiv.org/abs/2610.05923
Chat with Paper: https://academy.dair.ai/papers/vera-scaling-verifiable-environments-for-agentic-co-evolution-2610.05923
与论文对话:https://academy.dair.ai/papers/vera-scaling-verifiable-environments-for-agentic-co-evolution-2610.05923
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