Paper Reading Collection · by qqtang-code

论文精读合集 Paper Reading Collection

每篇论文一个"图文配合"的精读页:全部 Figure / Table 以 300 DPI 高清内嵌于对应讲解段落,配逐点解读、术语速查与编者点评。每篇均提供中文精读版英文版(两篇旗舰论文为完整英译,其余为英文速读)。 One illustrated reading page per paper: every figure and table embedded at 300 DPI in the paragraph that discusses it, with point-by-point commentary, a glossary and editorial notes. Every paper ships a full Chinese deep-read plus an English edition (complete translations for the two flagship papers, quick-read editions for the rest).
图文配合Illustrated 300 DPI 深色模式Dark mode KaTeX 打印友好Print-friendly 中英双语中文 · English
论文 66 papers
图表 46 图 + 5946 figures + 59 tables
分类 2 大类2 categories
语言 中 / ENLanguages ZH / EN
Category 01

⚡ 高效推理与部署 ⚡ Efficient Inference & Deployment

稀疏注意力 · KV Cache · 长上下文Sparse Attention · KV Cache · Long Context

① Declarative Attention
Language Models Can Control Their Own Attention
① Declarative Attention
Language Models Can Control Their Own Attention

让模型在思维链里用标签声明"自己要看哪里",引擎解析声明并跳过大部分 KV cache 读取。零训练零样本:注意力读取降 52.0% / 31.1%,精度仅降 1.27pp / 2.75pp,解码墙钟 0.71× / 0.77× The model declares where to look with tags in its chain-of-thought; the engine parses the declarations and skips most KV-cache reads. Zero training, zero-shot: attention reads down 52.0% / 31.1%, accuracy down only 1.27pp / 2.75pp, decode wall-time 0.71× / 0.77×.

2026-0910 图 10 表10 figures · 10 tablesarXiv:2609.02737 中文精读 + 英文完整版ZH deep-read + full EN mirror
MoE · 边缘部署 · 带宽自适应MoE · Edge Deployment · Bandwidth-Adaptive

② FreeToken
边缘原生的 MoE 服务系统
② FreeToken
Edge-Native MoE Serving with Bandwidth-Adaptive Execution

把个人电脑变成统一的弹性推理平台:带宽自适应执行让 8GB 笔记本跑 35B、台式机跑 284B、单张工作站 GPU 跑 753B GLM-5.2,快到能支撑真实 agent 负载。 Turns the consumer PC into one elastic inference platform: bandwidth-adaptive execution puts a 35B model on an 8GB laptop, 284B on a desktop, and 753B GLM-5.2 on a single workstation GPU — fast enough for real agent workloads.

2026-095 图 1 表5 figures · 1 tablearXiv:2608.16157 中文精读 + 英文速读ZH deep-read + EN quick-read
量化 · NVFP4 · 温度缩放Quantization · NVFP4 · Temperature Scaling

③ ReSET
Accurate Latency-Critical NVFP4 Reasoning
③ ReSET
Step-Aware Temperature Scaling for NVFP4 Reasoning

NVFP4 低精度推理在延迟敏感场景的精度救星:步骤感知温度缩放按解码步骤动态调整 softmax 温度,不牺牲延迟地恢复量化精度。 An accuracy rescue for latency-critical NVFP4 inference: step-aware temperature scaling adapts the softmax temperature per reasoning step — recovering quantized accuracy at zero weight cost, plus a CUDA-core small-M decode kernel.

2026-088 图 16 表8 figures · 16 tablesarXiv:2606.13233 中文精读 + 英文速读ZH deep-read + EN quick-read
KV Cache 驱逐 · 推理模型 · 高效服务KV Cache Eviction · Reasoning Models · Efficient Serving

④ Random Attention
重新思考面向高效推理的 KV Cache 驱逐
④ Random Attention
Rethinking KV Cache Eviction for Efficient Reasoning

KV cache 驱逐不需要打分:保住 prompt + 每个 KV 头内均匀随机驱逐,4 模型 × 6 任务追平最强基线、60 格中 31 格显著领先,vLLM 吞吐再快 32–43%——一篇把"选择信号几乎不贡献精度"钉死的机制论文。 KV cache eviction needs no score: pin the prompt and evict uniformly at random within each head — matching the strongest evictor across 4 models × 6 reasoning tasks (significantly ahead in 31 of 60 cells) while serving 32–43% higher throughput in vLLM, with a full mechanistic account of why the selection signal buys almost nothing.

2026-095 图 11 表5 figures · 11 tablesarXiv:2609.03430 中文精读 + 英文速读ZH deep-read + EN quick-read
KV Cache 压缩 · MoE · 多模态 · 长上下文KV Cache Compression · MoE · Multimodal · Long Context

⑥ DeepSeek-V4.1-Flash
552B 多模态 MoE 高效推理技术报告
⑥ DeepSeek-V4.1-Flash
Technical Report · 552B Multimodal MoE

DeepSeek-V4.1-Flash 技术报告:552B 多模态 MoE 用 Causal Encoder-Decoder 让 prefill 只激活 8B;CSA2 跨层 KV 复用 + FP4 量化把全局 KV Cache 压到每 token 890 字节(V4-Flash 的 1/4、V1 的 1/437),SWA Bounded Replay 再把持久缓存压到上代 1/8——性能反超更大的 V4-Flash:DeepSWE v1.1 74.2、Codeforces 3471 The DeepSeek-V4.1-Flash technical report: a 552B multimodal MoE whose Causal Encoder-Decoder activates just 8B in prefill; CSA2 cross-layer KV reuse + FP4 squeeze the global KV cache to 890 bytes/token (1/4 of V4-Flash, 1/437 of V1) while SWA Bounded Replay shrinks the persistent cache to 1/8 of the last generation — outperforming the larger V4-Flash: DeepSWE v1.1 74.2, Codeforces 3471.

2026-0912 图 5 表12 figures · 5 tablesGitHub 中文精读 + 英文完整版ZH deep-read + full EN mirror
Category 02

📏 评测基准 📏 Benchmarks

多模态 · 嵌入模型 · 长上下文Multimodal · Embeddings · Long Context

⑤ MMLongEmbed
长上下文场景下的多模态嵌入模型基准
⑤ MMLongEmbed
Benchmarking Multimodal Embedding Models in Long-Context Scenarios

首个专门评测多模态嵌入模型长上下文能力的基准:4 个检索任务、8 个数据集、8,460 个查询、11 个模型。核心发现:现役模型靠表面特征匹配"作弊",细粒度信息保持随长度衰退——"窗口大"≠"读得懂"。 The first benchmark dedicated to MEM long-context ability: 4 retrieval tasks, 8 datasets, 8,460 queries, 11 models. Key finding: current models "cheat" via surface-feature matching and fine-grained retention decays with length — a wide window ≠ real comprehension.

2026-086 图 16 表6 figures · 16 tablesGitHub 中文精读 + 英文速读ZH deep-read + EN quick-read