跳到主内容
@wquguru
精选88Rohan Paul论文研究

Xiaom发布MiMo-V2.6论文:AI自主训练循环与RL扩展

The paper for MiMo-V2.6 by Xiaom is out.

原文
发到 X
推荐理由

这篇论文揭示了AI自主闭环训练的新范式,特别是用评分代理解决代码RL中的投机问题,对做Agent训练的同学极具参考价值。

The paper for MiMo-V2.6 by Xiaom is out.

Xiaom 的 MiMo-V2.6 论文已发布。

In MiMo-V2.6, AI runs much of its own training loop: agents build the tasks, audit the tests, grade the answers and hunt for cheats. Humans set the budget and the rules.

在 MiMo-V2.6 中,AI 运行其大部分训练循环:智能体构建任务、审核测试、评估答案并搜寻作弊行为。人类设定预算和规则。

shows that agent models kept improving with more RL compute by scaling batch size, task and harness variety, and grading effort together.

表明随着批量大小、任务和工具链多样性以及评估工作量的扩展,智能体模型通过增加强化学习计算量持续改进。

Scaling RL for coding agents is hard because pass/fail tests can't tell a clean fix from a hacky fix. Agents also learn to game environments, for example by downloading the published fix.

为编码智能体扩展强化学习很困难,因为通过/失败测试无法区分干净的修复与 hacky 的修复。智能体还学会利用环境漏洞,例如下载已发布的修复补丁。

A grader agent compared passing patches in each group and moved reward to the cleaner ones. Without it, agents drifted toward longer runs and workarounds like swallowed exceptions.

一个评估智能体比较了各组中的通过补丁,并将奖励导向更干净的补丁。如果没有它,智能体会倾向于更长的运行时间和诸如吞没异常之类的变通方法。

MiMo-V2.6-Pro’s DeepSWE score rose from 58.4 to 72.6 over $2.6M of RL and was still climbing when training stopped.

MiMo-V2.6-Pro 的 DeepSWE 分数在 260 万美元的强化学习投入后从 58.4 升至 72.6,且在训练停止时仍在上升。

– arxiv. org/abs/2610.11959

– arxiv.org/abs/2610.11959

Title: "MiMo-V2.6: Scaling Reinforcement Learning Towards Self-Improvement"

标题:《MiMo-V2.6:扩展强化学习以实现自我改进》

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

进入量化体系 →

关联讨论

同一事件的更多信源

相似阅读

关联信息,但可能不是同一事件