固定模型改框架：编码Agent结果可大幅提升
Keep the model fixed, change the harness, and coding-agent results can move a lo…
Keep the model fixed, change the harness, and coding-agent results can move a lot when context gets tight.
保持模型不变，更换工作框架，当上下文变得紧张时，编码代理的结果可能会有很大波动。
The study compares 2 configurations of Yuj: control keeps the full chronological transcript until context fills, while treatment shortens older tool outputs, detects stalled behavior, and applies fixed command safeguards while preserving the full record.
该研究比较了Yuj的两种配置：对照组保留完整的按时间顺序的转录直到上下文填满，而处理组缩短较旧的工具输出，检测停滞行为，并在保留完整记录的同时应用固定的命令保障措施。
On 169 SWE-bench Verified tasks with a 20,480-token window, Qwen3.6’s mean per-task F2PF rose from 28% to 49%, while complete solutions increased from 43 to 72.
在169个SWE-bench Verified任务中，使用20,480个令牌的窗口，Qwen3.6的平均每任务F2PF从28%上升到49%，而完整解决方案从43个增加到72个。
The same frozen treatment improved both outcomes for Devstral, Nemotron, and Qwen3.8 without retuning.
相同的冻结处理在不重新调整的情况下，对Devstral、Nemotron和Qwen3.8的结果均有改善。
At 262,144 tokens, however, Verified and Pro outcomes were nearly identical between arms, so the benefit appears strongest when context is actually binding.
然而，在262,144个令牌时，Verified和Pro的结果在两组之间几乎相同，因此这种益处似乎在上下文真正受限时最为显著。
Treatment also used more model work under pressure, and the experiment tests the package as a whole rather than isolating each mechanism.
处理组在压力下也使用了更多的模型工作，该实验将整个包作为一个整体进行测试，而不是孤立每个机制。
For coding agents, “which model?” is no longer enough. Benchmark the model, harness, context policy, tools, and run controls as one solver.
对于编码代理来说，“用哪个模型？”已不再足够。将模型、工作框架、上下文策略、工具和运行控制作为一个整体求解器进行基准测试。
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