AI Research & Inference Systems
AI研究与推理系统
Akarsh Kumar, Phillip Isola
arXiv:2606.06479 · 2026-06-04T17:57:33Z
Supervised Memory Training uses a Transformer encoder to generate one-step memory-transition labels, then trains a nonlinear RNN without backpropagation through time. This creates a time-parallel training path with O(1) gradient length between tokens and outperforms BPTT in the reported language and pixel-sequence experiments. The preprint could make recurrent inference architectures easier to train, but broader model-scale and hardware-efficiency results are still needed.
Supervised Memory Training先用Transformer编码器生成单步记忆状态转移标签，再训练非线性RNN，从而避免时间反向传播。该方法形成可按时间并行的训练路径，使任意token间的梯度路径长度为O(1)，并在论文报告的语言与像素序列实验中优于BPTT。这项预印本可能降低循环推理架构的训练难度，但仍需更大模型规模与硬件效率数据。
Yutao Sun, Yanqi Zhang, Li Dong, et al.
arXiv:2606.06467 · 2026-06-04T17:54:04Z
Cross-Layer Sparse Attention computes token-level top-k routing once and reuses the index across decoder layers in a YOCO-style KV-sharing architecture. Sharing both KV state and routing attacks prefill cost, cache capacity, and decode overhead without falling back to coarse block sparsity. The paper reports quality across short- and long-context benchmarks, but the preprint still needs independent end-to-end serving validation.
Cross-Layer Sparse Attention在YOCO式KV共享架构中只计算一次token级top-k路由，并在多个解码层复用该索引。同时共享KV状态与路由，可共同降低prefill成本、cache容量和解码开销，又无需退回到较粗粒度的块稀疏。论文在短上下文与长上下文基准上报告了质量结果，但仍需独立的端到端服务验证。
Zhuoming Chen, Xinrui Zhong, Qilong Feng, et al.
arXiv:2606.06453 · 2026-06-04T17:48:17Z
Vortex combines a Python-embedded sparse-attention language with a page-centric tensor abstraction and a backend integrated into LLM serving stacks. Algorithms generated and refined through the system reach up to 3.46x the throughput of full attention while preserving accuracy in the reported tests. The useful contribution is programmability plus deployment, although the best results depend on the paper's benchmark and serving environment.
Vortex将嵌入Python的稀疏注意力描述语言、以页面为中心的张量抽象，以及集成到LLM服务栈的后端结合起来。通过该系统生成并优化的算法，在论文测试中可达到完整注意力最高3.46倍的吞吐量，同时保持准确率。其关键贡献是把可编程性与部署结合，但最佳结果仍依赖论文所用基准和服务环境。
Yasmine Omri, Ziyu Gan, Zachary Broveak, et al.
arXiv:2606.06448 · 2026-06-04T17:44:18Z
This study profiles ten agent-memory systems across two benchmark suites and separates construction, retrieval, and generation costs. Its four-axis taxonomy and ten system recommendations expose how designs move work between read and write paths, including scheduling, freshness, and amortization tradeoffs. The characterization is useful for storage-tier and fleet planning, but conclusions may shift as agent workloads and memory implementations mature.
该研究在两套基准上分析了10种Agent记忆系统，并将构建、检索与生成成本分别归因。其四维分类法与10项系统建议揭示了不同设计如何在读写路径之间转移工作，包括调度、新鲜度与成本摊销权衡。这些结果有助于存储分层与集群规划，但随着Agent负载和记忆实现演进，结论仍可能变化。
Hyungmin Kim, Minsoo Kim, Hongseok Kim, et al.
arXiv:2606.06302 · 2026-06-04T15:41:27Z
Tangram assigns a deterministic memory budget to each attention head, groups heads with similar retention needs into vectorized page tables, and balances work ahead of time. These choices target the fragmentation, scheduling, and kernel-utilization penalties that have limited non-uniform KV-cache compression in multi-turn serving. The architecture is well matched to GPU memory constraints, though the available abstract does not expose enough end-to-end numbers to judge deployment gains.
Tangram为每个注意力头分配确定性的内存预算，将保留需求相近的头组织进向量化页表，并提前完成负载均衡。这些设计针对多轮服务中限制非均匀KV cache压缩的内存碎片、调度复杂度与内核利用率损失。其架构与GPU内存约束高度相关，但现有摘要缺少足够的端到端数据，尚难判断实际部署收益。