Athanor：面向自主后训练研究的智能体原生系统

Athanor: An Agent-Native System for Autonomous Post-Training Research

arXiv ID: arXiv:2610.21055 [cs.AI]
Submitted On: October 15, 2026
Primary Subject: Artificial Intelligence (cs.AI)
Other Subjects: Computation and Language (cs.CL), Machine Learning (cs.LG), Multiagent Systems (cs.MA)

文章背景与核心概要

如果让 AI 智能体自己动手训练大模型，它们能做到什么程度？来自多所顶尖研究机构的学者们推出了名为 Athanor 的开源智能体原生研究系统。在完全脱离人类干预且无法调用外部更强“教师模型”的严苛沙盒环境中，四名由 Claude Opus 5.5 驱动的 AI 智能体在真实 GPU 集群上展开了一场长达 104 小时的自主科研战役。它们不仅独立构建数据集、编写训练算法并调优超参数，还灵活运用了强化学习、自蒸馏与权重融合技术。最终，智能体团队成功将原生开源基座模型 Qwen3-4B-Base 锻造为全能模型，其跨领域综合测试表现甚至一举超越了官方发布的同规模指令微调版本，首次有力证明了前沿智能体已具备自主驱动深度学习科研与模型进化的硬核工程能力。

作者团队

Authors

Yufa Zhou
Yichen Wang
Jiakun Fan
Ziqi Shu
Zhichao Liu
Anru R. Zhang
Neiwen Ling
Zijun Wei

Yufa Zhou

Yichen Wang

Jiakun Fan

Ziqi Shu

Zhichao Liu

Anru R. Zhang

Neiwen Ling

Zijun Wei

执行摘要

Executive Summary

Athanor 是一套开源的智能体原生系统，旨在通过真实图形处理器 (GPU) 上的自主后训练研究，深入探索 AI 系统的递归自改进 (Recursive Self-Improvement) 机制。在完全脱离外部教师模型指导且零人类干预的严格沙盒环境中，一个由四名 Claude Opus 5.5 驱动的 AI 智能体 (AI Agent) 团队，成功将未经调优的原始基座模型 Qwen3-4B-Base 锤炼成了一款跨领域通用大语言模型 (Large Language Model, LLM)。在长达 104 小时的持续科研推进中，智能体团队自主清洗与构建数据集、编写全新训练算法、调优超参数，并创造性地融合了自蒸馏 (Self-Distillation) 与权重融合 (Weight Merging) 技术。

Athanor is an open, agent-native system designed to study recursive self-improvement through autonomous post-training research on real GPUs. Operating in a strict sandbox without external teacher models or human intervention, a team of four Claude Opus 5.5 agents successfully transformed the raw base model Qwen3-4B-Base into a cross-domain generalist model. Over a 104-hour campaign, the agents built datasets, wrote training algorithms, tuned hyperparameters, and applied self-distillation and weight merging.

最终训练出的模型展现出极其惊人的性能飞跃：

The resulting model improved dramatically:

开发测试集 (包含 6 项核心任务)： 综合评分从 15.9 跃升至 67.3 (相较之下，官方指令微调版 Qwen3-4B-Instruct-2507 得分为 66.5)；
独立保留测试集 (Held-out Suite)： 评分从 9.2 飙升至 48.2 (甚至超过了官方发布的指令微调版本的 47.9)。

Development Suite (6 tasks): Increased from 15.9 to 67.3 (compared to Qwen3-4B-Instruct-2507 at 66.5).

Held-out Suite: Increased from 9.2 to 48.2 (compared to the official instruct version at 47.9).

据我们所知，这是业内首次成功验证：前沿 AI 智能体完全依靠自主后训练，便能将未经打磨的原始基座模型提升至匹敌甚至超越官方精调指令模型的水平。

To our knowledge, this is the first successful demonstration of frontier agents autonomously post-training a raw base model to match or surpass the performance of an official instruct version.

论文摘要

Abstract

递归自改进——即让 AI 系统自主构建更强大的下一代 AI 系统——一直被视作通往超级智能 (Superintelligence) 的极具前景的路径之一。在本项研究中，我们聚焦于一种在当下具备可测性与可落地性的探索范式：通过让前沿智能体仅利用公开数据集对小型原始基座模型进行后训练，来全方位检验其自主科研潜力与工程极限。

Recursive self-improvement, in which AI systems build stronger AI systems, is one promising route to superintelligence. We study a weaker form that can be measured today: frontier agents post-training a small raw base model with public data, as a test of their research potential.

为此，我们构建了 Athanor——一套面向真实 GPU 环境下自主后训练研究 (Autonomous Post-Training Research, AutoPostTrain) 的开源智能体原生系统。我们赋予了智能体完全的科研自主权、低门槛的顺畅工程工具链以及高效率的初始脚手架代码。整个科研探索完全由智能体自主驱动：它们独立查阅数据集与评测反馈、自主构建微调样本、编写并迭代训练算法、调优关键超参数，并严谨分析实验日志以决定下一阶段的探索方向。所有智能体均在安全沙盒中独立运作，且严禁调用外部更强大的教师模型。

We build Athanor, an open agent-native system for autonomous post-training research (AutoPostTrain) on real GPUs. We give the agents full research autonomy, low-friction engineering tools, and efficient starter code. The agents carry out the research themselves: they read datasets and evaluation results, build their training data, write and modify training algorithms, tune hyperparameters, and analyze the results to decide what to try next. They work in a sandbox and cannot use external teacher models.

Athanor 的后台服务提供了固化的评测基准、托管式作业调度、严格去污染的数据集、持久化科研日志，以及智能体事先毫不知情的盲测独立保留测试集。在一场耗时 104 小时的连续实验中，由四名各自独占 8 卡 GPU 算力节点的 Claude Opus 5.5 智能体组成的研究团队，在未获得任何外部人类科研建议的条件下，将 Qwen3-4B-Base 在包含 6 项任务的开发基准上的得分从 15.9 提升至 67.3，并在完全盲测的独立保留测试集上从 9.2 跃升至 48.2；作为对比，官方发布的 Qwen3-4B-Instruct-2507 在这两个测试集上的得分分别为 66.5 与 47.9。

Athanor's services supply frozen evaluation, managed jobs, decontaminated data, a persistent research record, and a held-out suite the agents are unaware of. In a 104-hour campaign, a team of four Claude Opus 5.5 agents, each on an eight-GPU node and given no research advice, improves Qwen3-4B-Base from 15.9 to 67.3 on a six-task development suite and from 9.2 to 48.2 on a held-out suite, against 66.5 and 47.9 for the official Qwen3-4B-Instruct-2507.

实验的完整科研轨迹记录表明，智能体之所以能取得这一卓越水准，关键在于它们学会了将已知的前沿算法与团队彼此的模型成果进行有机重组：包括在公开推理数据上执行监督微调 (Supervised Fine-Tuning, SFT)、应用某位智能体率先提出的采样展开改进版强化学习 (Reinforcement Learning, RL)、跨智能体间的权重融合，以及多轮深度自蒸馏。这表明只要底层的工程调度与环境交互问题得到有效解决，当下的 AI 智能体已然具备攻克复杂深度学习研发任务的硬核工程与科研能力。我们期待 Athanor 能为进一步扩展自主 AI 科研 (Autonomous AI Research) 的疆界奠定坚实基石。

The campaign's full record shows that the agents reach this level by recombining known methods and each other's models: SFT on public reasoning data, RL with a rollout change one agent introduced, weight merges across agents, and rounds of self-distillation. With execution handled, today's agents already have the engineering and research ability to conquer this task. We hope Athanor provides a foundation for scaling autonomous AI research.