视频选题榜 2026-08-12

• https://arxiv.org/abs/2608.08311 · Ouroboros: A Self-Developing Frontier Coding Agent with Reviewed Core Evolution · 自演化 Agent harness 用 review commit 治理 · round=R1
• https://arxiv.org/abs/2608.09867 · Stealing Reasoning Traces from Proprietary LLM APIs · 加密推理块跨模型可互换=弱模型可代理解码强模型推理 · round=R2
• https://arxiv.org/abs/2608.11205 · AdvFD: Boosting Visual Generation via Adversarial Fréchet Distance Loss · 对抗式可学习表征+白化跳出视觉生成「指标涨=视觉质量垮」的 Fréchet hacking 陷阱 · round=R3