Agora: Git as Shared Memory for Collective AutoResearch

ID: 6c0a71a5
原文链接: https://arxiv.org/abs/2609.18094
作者: Yifan Zhang, Yunheng Zou, Shaokun Zhang, Jian Hu, Hao Zhang, Binfeng Xu, Jan Kautz, Yi Dong
日期: 2026-09-16
分类: auto
来源类型: paper
标签: arxiv, paper, multi-agent, autoresearch, shared-memory, weight-transfer
质量评分: 5/5
抓取时间: 2026-09-20T23:30Z

中文摘要

Agora 是 NVIDIA 在 2026-09-16 公开的 arXiv 技术报告（2609.18094，cs.LG/cs.AI/cs.CL）：把多 agent 协作的共享记忆写成 Git 里的 append-only DAG，每条贡献是不可变 commit + parent edges；用 derived index 暴露 frontier、neglected branches、verification status，再用 diversity-aware selection rule 防止社区塌到一个 leader 上。第一次持续运行是 12 天：13 个 LLM worker（Claude Opus 4.7 + GPT-5.5），无指派任务、无中央 planner；141 个 pretrained donor 初始化一个 frozen 119.6M attention-SSM hybrid，最终 1,703 个贡献把 evaluator 从 3.39 bpb 拉到 1.899 bpb，关闭到 GPT-2 124M 训练差距的 62%；winner recipe 的 145-commit ancestry 跨 15 个账号，165 次独立复现全通过。

为什么值得关注

把多 agent 共享记忆写成 Git DAG：13 个 worker 12 天跑出 1.703 个 commit，把 weight transfer 从 3.39 bpb 拉到 1.899 bpb。

English Abstract

Agora records multi-agent research contributions as an append-only DAG in Git, with each result/insight/hypothesis/verification/report as an immutable commit and parent edges describing what it builds on; a derived index exposes frontier, neglected branches, and verification status, while a diversity-aware selection rule prevents collapse onto one leader. First sustained run lasted nearly 12 days with 13 LLM workers (Claude Opus 4.7 / GPT-5.5), no assigned tasks and no central planner, working a weight-transfer problem over 141 pretrained donor models into a frozen 119.6M attention-SSM hybrid; the 1,703 contributions drove the evaluator from 3.39 to 1.899 bpb, closing 62% of the gap to a trained GPT-2 124M, with the winning recipe's 145-commit ancestry spanning 15 accounts and 165 independent reproductions all passing.

Obsidian 证据摘要

来源: OpenClaw定时任务/DevRadar + 调研/2026-09-20-调研-Agora多Agent工作地图.md