#代码理解

This research reveals that LLM performance degrades significantly as code complexity increases, a critical insight masked by aggregate accuracy metrics. Developers and organizations relying on AI for code review or debugging should consider complexity-specific performance when selecting models, as even advanced models like DeepSeek-Coder-V2 struggle with highly complex code structures.

研究SWE Agent行为的团队终于有了可量化的观察方法——Ada框架把黑箱行为变成可比较的轨迹画像，做Agent评估和调试的开发者值得关注。