We plugged Google Research’s Harness² into JEV 1.14. Zero weights touched, yet driving massive gains on vision benchmarks round after round. Under the hood, the system: ⚡️ Rewrites its own prompts 🛠️ Builds its own tools 🧠 Distills its own experience JEV-Harness² repo drops soon. Stay tuned.

Iterating fast, open source soon. Follow along🏊

I like this ui pipeline

The UI pipeline in this video will be open-sourced together with the repo.

这么好的内容为什么没人评论 哈哈

技术内容流量是低呀哈哈

所以我说我在看x 纯是实力淘金 😂

@bot add this info

零权重调整还能一轮轮涨，这个点很关键：说明瓶颈不一定在模型本身，而在提示、工具和经验如何被组织。想追问一下，这种自改进在未见过的视觉任务上泛化如何，会不会只对已见分布的错误模式有效？

someone has to read the prompts it writes

prompt rewrites run through propose-probe-compose, which proposes harness edits, probes them on the current task, and composes the next harness

Will the release include a fixed-harness baseline with the same total inference budget? That would help separate the benefit of the evolving harness from simply spending more compute per example.

Of course, that includes the data for the entire benchmark, which will be stored in the repository.