约翰霍普金斯与卡内基梅隆大学论文:小模型可自动优化 Agent 代码
New John Hopkins and Carnegie Mellon University Paper Shows that a small model c…
Agent 工程的核心痛点是调试,这篇论文给出了一种用更小的模型自动优化 Agent 框架代码的方法,且效果超越大模型教师,对做 Agent 的同学很有参考价值。
New John Hopkins and Carnegie Mellon University Paper Shows that a small model can learn to improve an agent's harness code from run results, and the skill transfers to new tasks.
约翰·霍普金斯大学和卡内基梅隆大学的新论文表明,一个小模型可以从运行结果中学习如何改进智能体的 harness 代码,并且该技能可以迁移到新任务中。
A small model trained to rewrite an agent's harness code from failure reports can adapt agents to new tasks, so let it tune your harness instead of doing it by hand.
一个根据失败报告重写智能体 harness 代码的小模型,可以使智能体适应新任务,因此让它来调整你的 harness,而不是手动操作。
The harness is the code that decides what the model sees and which tools it calls. The editor reads the harness and what failed, then writes a code change, rewarded by how well the new harness scores.
Harness 是决定模型看到什么以及调用哪些工具的代码。编辑器读取 harness 和失败原因,然后编写代码更改,其奖励取决于新 harness 的评分表现。
On 21 unseen reasoning task types, a 4B editor's average edit score rose from 0.32 to 0.62, above its 35B teacher. A separate editor trained on HotpotQA kept improving harnesses on 2 other QA benchmarks.
在 21 种未见过的推理任务类型上,4B 编辑器的平均编辑得分从 0.32 提升至 0.62,超过了其 35B 教师模型的水平。另一个在 HotpotQA 上训练的独立编辑器在其他两个问答基准测试上持续改进了 harness。
Run it for several rounds per task, since repeated edits beat single fixes.
每个任务运行几轮,因为多次编辑优于单次修复。
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