视频选题榜 2026-08-20

• https://arxiv.org/abs/2608.17950 · Do Large Language Models Play Six Degrees of Separation? Measuring Topological Compression in Long-Context Manifolds · deep LLM 潜空间天然组织为 Small-World 网络 + 任意语义锚点 ≤6 跳可达 + 事实正确 ≈3 跳结构连通 + 拓扑塌缩 = 幻觉零样本检测 · round=R1
• https://arxiv.org/abs/2608.14376 · CoRun: Deterministic LLM Inference Scheduling · continuous batching 形状固定 + prefill 隔离 + decode pad + 固定 CUDA Graph replay + 吞吐量 +15-324%(>2×) + 输出 bit-identical · round=R2
• https://arxiv.org/abs/2608.14333 · DASH: HBM+Flash KV Cache Management for MoE LLM · MoE LLM KV-cache HBM+Flash 两级管理 + HBM 吸收细粒度写入 + 批量写回 + Llama 4 Maverick 197.41GB 上 1.92× 吞吐 + 48% 延迟 · round=R2
• https://arxiv.org/abs/2608.18701 · SoftVTBench: A Deformation-Aware Visuo-Tactile Dataset and Benchmark for Deformable-Object Manipulation · 可变形物体操作的安全感知视-触觉基准 · Goal Success + Safety Success 双层指标 + 4,000 demonstrations + 50+ assets + 20 Hz + Isaac Sim + 有限元 FEM · round=R3