🔥 毒舌 GitHub
← 返回论文榜
Understanding by Reconstruction: Reversing the Software Development Process for LLM Pretraining
Zhiyuan Zeng, Yichi Zhang, Yong Shan, Kai Hua et al.
39.60/100
💧
灌水嫌疑 · 价值存疑
内容分 39.6 · 引用加成 +0.0 · 暂无引用数据

💡 本文提出通过多智能体模拟反向合成软件开发中间轨迹用于LLM预训练,宣称可提升模型长程推理与代码能力

#代码预训练画饼#轨迹合成#多智能体套娃#长程推理玄学#实验摆烂#code pre-training pie-in#trajectory synthesis#multi-agent nesting#long-horizon reasoning m#shoddy experiments

维度评分

创新性6.0 / 10
严谨性3.0 / 10
意义6.0 / 10
清晰度7.0 / 10
可复现性3.0 / 10

这个语气还没生成 —— 去重新评测一次 即可生成。