视频选题榜 2026-09-09

• https://arxiv.org/abs/2609.04010 · Unlocking Lossless Speedups in LLMs via Discrete Diffusion · 保留 AR 分布并用轻量扩散头并行采样多个 token · round=R1
• https://arxiv.org/abs/2609.04971 · BeaconKV: Key-Value Cache Compression Guided by Beacon Queries for Efficient Large Reasoning Model Inference · 长链推理显存瓶颈,beacon query 引导 KV 压缩,5.8× 显存 + 4.3× 吞吐 · round=R3