视频选题榜 2026-07-14

• https://arxiv.org/abs/2603.15569 · Mamba-3: Improved Sequence Modeling using State Space Principles · 三项正交改进（trapezoidal + complex + MIMO）推进次二次 Pareto 前沿 · round=R2
• https://arxiv.org/abs/2604.12374 · Nemotron 3 Super: Open, Efficient Mixture-of-Experts Hybrid Mamba-Transformer Model for Agentic Reasoning · 120B/12B active hybrid MoE + 1M context + PinchBench 85.6% · round=R2（备选 3）
• https://arxiv.org/abs/2604.14191 · Attention to Mamba: A Recipe for Cross-Architecture Distillation · Apple 两阶段蒸馏（Transformer → linearized → Mamba）· 14.11 vs 13.86 Pythia-1B perplexity · round=R2（备选 5）
• https://arxiv.org/abs/2607.09424 · Soofi S: Sovereign German-English Foundation Model · KI Bundesverband + Deutsche Telekom 30B-A3B hybrid Mamba-Transformer MoE · 1M context · round=R2（备选 4）- https://arxiv.org/abs/2602.06036 · DFlash: Block Diffusion for Flash Speculative Decoding · HyOCR-1.5 引入的投机解码框架（Transformers 6.37× / vLLM 2.14×）· round=R3