视频选题榜 2026-07-22

• https://arxiv.org/abs/2607.18171 · FlashRT: Agent Harness for Guiding Agents to Deploy Real-Time Multimodal Applications · 让通用 coding agent 自动把多模态参考实现转成多 GPU 高效部署 — placement + streaming + intra-model parallelism 三件套解耦为多轮 Agent 变换 · round=R1
• https://arxiv.org/abs/2607.18213 · SWE-Pruner Pro: The Coder LLM Already Knows What to Prune · 在 Agent 内部挂 keep-or-prune 头按行裁剪 tool 输出 — 最多省 39% token + SWE-Bench Verified resolve rate 反向 +3.8% · round=R1
• https://arxiv.org/abs/2607.18217 · HOMIE: Human-object Centric Video Personalization via Multimodal Intelligent Enhancement · MLLM 整合 + 全局多模态引导 + 模态参考嵌入 — 在多种 HOCVP 任务上达 SOTA · round=R1
• https://arxiv.org/abs/2607.18754 · AgentDebugX: An Open-Source Toolkit for Failure Observability, Attribution, and Recovery in LLM Agents · 把 LLM Agent 调试组织成 Detect→Attribute→Recover→Rerun 闭环 + DeepDebug 多轮根因诊断 — Who&When qwen3.5-9b 28.8% 严格归因 + GAIA 单次 rerun 修复 13/73 + 整体准确率 55.8%→63.6% · round=R3