清华等高校论文：AI Agent 训练易陷入局部优化，缺乏策略重构机制
New Tsinghua and other Chinese Univ paper finds AI agents can optimize a trainin…
直击当前 AutoML 与 AI R&D Agent 的核心痛点：缺乏元认知层面的策略重构能力。这篇实证研究为构建更自主的训练代理提供了关键的设计启示，值得 Agent 开发者深入参考。
New Tsinghua and other Chinese Univ paper finds AI agents can optimize a training plan for hours, but rarely realize when the plan itself is wrong.
清华大学及其他中国高校的新论文发现，AI 智能体可以优化训练计划长达数小时，但很少能意识到计划本身是错误的。
Once training starts, an AI agent's first strategy tends to become the strategy it keeps optimizing.
一旦训练开始，AI 智能体的首要策略往往会成为它持续优化的策略。
Across 1,338 post-training trajectories, only 74 of 3,557 adjacent training experiments, 2.1%, changed the high-level strategy. Most iterations stayed inside the same approach, tweaking data, hyperparameters, formatting, or bugs.
在 1,338 条后训练轨迹中，3,557 组相邻训练实验仅有 74 组（2.1%）改变了高层策略。大多数迭代都停留在同一方法内，仅调整数据、超参数、格式或修复 bug。
The agents were not inactive. They could train, evaluate, diagnose failures, and improve checkpoints.
这些智能体并非无所作为。它们能够进行训练、评估、诊断失败原因并改进检查点。
The problem appeared when evidence suggested the whole approach should change.
当证据表明整个方法需要改变时，问题便显现出来。
Adding an experiment journal, skill library, and evaluator lifted GSM8K by 12.6 points and HumanEval by 40.8 points over the Opus 4.6 baseline, but still did not trigger strategy changes.
增加实验日志、技能库和评估器使 GSM8K 提升了 12.6 分，HumanEval 提升了 40.8 分（相对于 Opus 4.6 基线），但仍未能触发策略变更。
Human guidance could redirect the plan before training, yet the agent returned to local tuning once training began.
人类指导可以在训练前重新规划方向，但一旦训练开始，智能体又回到了局部微调。
Even 2–8× more inference tokens mostly bought more refinement, with almost no reliable gain on AIME 2025.
即使推理 token 数量增加 2–8 倍，主要带来的也只是更细致的优化，在 AIME 2025 上几乎没有带来可靠的提升。
For AI R&D agents, the missing mechanism may be an explicit trigger to reopen the strategy itself, not just keep optimizing it.
对于 AI 研发智能体而言，缺失的机制可能是明确触发重新审视策略本身的机制，而不仅仅是继续优化它。
– arxiv. org/abs/2608.19072
– arxiv.org/abs/2608.19072
Title: "What is Missing from AI Post-Training AI: An Empirical Analysis"
标题：《AI 后训练 AI 缺了什么：一项实证分析》
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