{ "snapshot_date": "2026-10-20", "scope": "local library; not a global publication census", "counts": { "C": 67, "B": 23, "A": 45 }, "groups": [ { "id": "capacity", "zh": "扩容与多记忆", "en": "Capacity / multiple memories", "members": [ "Triadic", "SDM", "MoM", "FwPKM", "MDN", "DSLA" ] }, { "id": "geometry", "zh": "写入、读取与容量调度", "en": "Write / read / allocation", "members": [ "Lattice", "PDN", "OSDN", "KDN", "BayesianLayer", "CCQ", "EDA", "Q-Delta", "MesaNet", "Proteus", "CyFA" ] }, { "id": "dynamics", "zh": "时间尺度与状态变换", "en": "Timescales / transitions", "members": [ "MS-GLA", "FRAC", "DeltaProduct", "CKDA", "SFDA", "SelectiveRoPE", "Mamba-3", "Mamba3-Oscillation", "M2RNN", "SpectralShift", "PaTH" ] }, { "id": "kernels", "zh": "核与压缩表示", "en": "Kernels / representations", "members": [ "BASED", "LAVO", "RACE", "SANATAN" ] }, { "id": "hybrids", "zh": "压缩记忆与显式检索组合", "en": "Compressed / explicit hybrids", "members": [ "LoLA", "HOLA", "NAtS-L", "KVM", "MiTA", "Samba", "YOCO", "Nectar" ] }, { "id": "training", "zh": "训练、转换与长度泛化", "en": "Training / conversion", "members": [ "JRT", "LoLCATs", "HedgeCATs", "HypeNet", "HybridPosition", "OAL", "LengthGen", "DAT" ] }, { "id": "systems", "zh": "状态压缩与服务系统", "en": "State storage / serving", "members": [ "StateReduction", "DAMP", "LeapQuant", "DeltaLog", "HYPIC", "Tail-Replay" ] }, { "id": "analysis", "zh": "机制、理论与综述", "en": "Mechanisms / theory / reviews", "members": [ "LinearAttn-Memory", "CLVR", "LLM-Memory", "Attention-Evolution", "SSM-Gating", "Hybrid-Hierarchy", "Transformer-Topology" ] }, { "id": "fastmemory", "zh": "快权重与神经记忆近邻", "en": "Fast weights / nearby memory", "members": [ "TTT", "LaCT", "SRWM", "NL", "MoNe", "MemoryMosaics", "Metis" ] } ], "papers": [ { "name": "AlgebraLayers", "directory": "AlgebraLayers-2609.32814", "title": "Change the Product, Keep the Parameters: Associative Algebra Layers for Transformers", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/AlgebraLayers-2609.32814/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/AlgebraLayers-2609.32814/v1/source/paper.pdf", "source_directory": "paper-readings/AlgebraLayers-2609.32814/v1/source", "source_url": "https://arxiv.org/abs/2609.32814v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Attention-Evolution", "directory": "Attention-Evolution-2609.39661", "title": "The Evolution of Attention in Large Language Models: Mechanisms, Trade-offs, and Emerging Trends", "version": "v1", "revision": "r1", "tier": "B", "primary_category": "analysis", "assessment": "从表示到整合梳理注意力和混合架构；说明研究设计轴众多，不计作一种独立更新算法。", "notes": "paper-readings/Attention-Evolution-2609.39661/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/Attention-Evolution-2609.39661/v1/source/paper.pdf", "source_directory": "paper-readings/Attention-Evolution-2609.39661/v1/source", "source_url": "https://arxiv.org/abs/2609.39661v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Attention-Temperature", "directory": "Attention-Temperature-2505.14840", "title": "Subquadratic Algorithms and Hardness for Attention with Any Temperature", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/Attention-Temperature-2505.14840/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/Attention-Temperature-2505.14840/v1/source/paper.pdf", "source_directory": "paper-readings/Attention-Temperature-2505.14840/v1/source", "source_url": "https://arxiv.org/abs/2505.14840v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "BASED", "directory": "BASED-2402.18668", "title": "Simple linear attention language models balance the recall-throughput tradeoff", "version": "v2", "revision": "r1", "tier": "A", "primary_category": "kernels", "assessment": "二阶 Taylor 全局状态、小窗口精确注意力及卷积组合；将容量、局部精度和 IO 内核一起设计。", "notes": "paper-readings/BASED-2402.18668/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/BASED-2402.18668/v2/source/paper.pdf", "source_directory": "paper-readings/BASED-2402.18668/v2/source", "source_url": "https://arxiv.org/abs/2402.18668v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "BayesianLayer", "directory": "BayesianLayer-2605.31163", "title": "Memory by Design: Probabilistic Sequence Layers", "version": "v3", "revision": "r1", "tier": "A", "primary_category": "geometry", "assessment": "稠密协方差引导抗碰撞写入；增加二次辅助状态，召回改善伴随整体困惑度取舍。", "notes": "paper-readings/BayesianLayer-2605.31163/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/BayesianLayer-2605.31163/v3/source/paper.pdf", "source_directory": "paper-readings/BayesianLayer-2605.31163/v3/source", "source_url": "https://arxiv.org/abs/2605.31163v3", "available_source_versions": 2, "available_note_revisions": 2 }, { "name": "BinaryPC", "directory": "BinaryPC-2608.04405", "title": "Training-Free Hashing-Based Attention via Binary Principal Components", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/BinaryPC-2608.04405/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/BinaryPC-2608.04405/v1/source/paper.pdf", "source_directory": "paper-readings/BinaryPC-2608.04405/v1/source", "source_url": "https://arxiv.org/abs/2608.04405v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "BoV", "directory": "BoV-2606.02780", "title": "Do Value Vectors in Deep Layers Need Context from the Residual Stream?", "version": "v5", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/BoV-2606.02780/v5/r1/reading-notes.zh.md", "pdf": "paper-readings/BoV-2606.02780/v5/source/paper.pdf", "source_directory": "paper-readings/BoV-2606.02780/v5/source", "source_url": "https://arxiv.org/abs/2606.02780v5", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "CALM", "directory": "CALM-2510.27688", "title": "Continuous Autoregressive Language Models", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/CALM-2510.27688/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/CALM-2510.27688/v1/source/paper.pdf", "source_directory": "paper-readings/CALM-2510.27688/v1/source", "source_url": "https://arxiv.org/abs/2510.27688v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "CCQ", "directory": "CCQ-2606.01294", "title": "Don't Read Everything: A Curvature-Conditioned Query for Linear Attention", "version": "v2", "revision": "r1", "tier": "A", "primary_category": "geometry", "assessment": "以历史协方差收缩 query；新增矩阵和均值，且关闭推理修正仍保留多数收益，需区分训练作用。", "notes": "paper-readings/CCQ-2606.01294/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/CCQ-2606.01294/v2/source/paper.pdf", "source_directory": "paper-readings/CCQ-2606.01294/v2/source", "source_url": "https://arxiv.org/abs/2606.01294v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "CDLM", "directory": "CDLM-2605.00161", "title": "Consistent Diffusion Language Models", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/CDLM-2605.00161/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/CDLM-2605.00161/v2/source/paper.pdf", "source_directory": "paper-readings/CDLM-2605.00161/v2/source", "source_url": "https://arxiv.org/abs/2605.00161v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "CKA", "directory": "CKA-1905.00414", "title": "Similarity of Neural Network Representations Revisited", "version": "v4", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/CKA-1905.00414/v4/r1/reading-notes.zh.md", "pdf": "paper-readings/CKA-1905.00414/v4/source/paper.pdf", "source_directory": "paper-readings/CKA-1905.00414/v4/source", "source_url": "https://arxiv.org/abs/1905.00414v4", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "CKDA", "directory": "CKDA-2609.24797", "title": "Complex KDA: Understanding and Enhancing the Expressivity of Kimi Delta Attention", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "dynamics", "assessment": "有符号通道门和扩展更新率以反射组合旋转；实数状态下增强追踪，LM 与匹配 KDA 基本持平。", "notes": "paper-readings/CKDA-2609.24797/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/CKDA-2609.24797/v1/source/paper.pdf", "source_directory": "paper-readings/CKDA-2609.24797/v1/source", "source_url": "https://arxiv.org/abs/2609.24797v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "CL-Composition", "directory": "CL-Composition-2609.06986", "title": "Continual Learning Mechanisms Compose for Long-Horizon Memorization", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "agent-memory", "assessment": "外部文档、智能体、参数记忆、数据或提示组织，未研究本报告的序列内循环 state", "notes": "paper-readings/CL-Composition-2609.06986/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/CL-Composition-2609.06986/v1/source/paper.pdf", "source_directory": "paper-readings/CL-Composition-2609.06986/v1/source", "source_url": "https://arxiv.org/abs/2609.06986v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "CLM", "directory": "CLM-2609.37725", "title": "Context Language Models", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "agent-memory", "assessment": "外部文档、智能体、参数记忆、数据或提示组织，未研究本报告的序列内循环 state", "notes": "paper-readings/CLM-2609.37725/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/CLM-2609.37725/v1/source/paper.pdf", "source_directory": "paper-readings/CLM-2609.37725/v1/source", "source_url": "https://arxiv.org/abs/2609.37725v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "CLVR", "directory": "CLVR-2607.07953", "title": "Linear Attention Architectures: Mechanisms, Trade-offs, and Cross-Layer Routing", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "analysis", "assessment": "统一遗忘/擦除/写入比较，研究跨层 value 路由；单次小幅损失改善尚非独立的容量扩展证据。", "notes": "paper-readings/CLVR-2607.07953/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/CLVR-2607.07953/v1/source/paper.pdf", "source_directory": "paper-readings/CLVR-2607.07953/v1/source", "source_url": "https://arxiv.org/abs/2607.07953v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "ClockDiffusion", "directory": "ClockDiffusion-2610.00894", "title": "Clock Diffusion: Efficient Semi-Autoregressive Continuous Diffusion Language Models", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/ClockDiffusion-2610.00894/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/ClockDiffusion-2610.00894/v1/source/paper.pdf", "source_directory": "paper-readings/ClockDiffusion-2610.00894/v1/source", "source_url": "https://arxiv.org/abs/2610.00894v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Cogentic", "directory": "Cogentic-2609.40324", "title": "Cogentic: Multi-Agent Orchestration for Automated Proof Discovery", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/Cogentic-2609.40324/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/Cogentic-2609.40324/v1/source/paper.pdf", "source_directory": "paper-readings/Cogentic-2609.40324/v1/source", "source_url": "https://arxiv.org/abs/2609.40324v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "ConceptLM", "directory": "ConceptLM-2602.08984", "title": "Next Concept Prediction in Discrete Latent Space Leads to Stronger Language Models", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/ConceptLM-2602.08984/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/ConceptLM-2602.08984/v1/source/paper.pdf", "source_directory": "paper-readings/ConceptLM-2602.08984/v1/source", "source_url": "https://arxiv.org/abs/2602.08984v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "CyFA", "directory": "CyFA-2609.36259", "title": "CyFA: Linear Sequence Modeling with Relative-Time-Partitioned Memory", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "geometry", "assessment": "用可学习时钟组织相对时间槽，双状态经精确坐标变换实现；主状态匹配下改善召回。", "notes": "paper-readings/CyFA-2609.36259/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/CyFA-2609.36259/v1/source/paper.pdf", "source_directory": "paper-readings/CyFA-2609.36259/v1/source", "source_url": "https://arxiv.org/abs/2609.36259v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "DA", "directory": "DA-2609.02737", "title": "Language Models Can Control Their Own Attention", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/DA-2609.02737/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/DA-2609.02737/v1/source/paper.pdf", "source_directory": "paper-readings/DA-2609.02737/v1/source", "source_url": "https://arxiv.org/abs/2609.02737v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "DAMP", "directory": "DAMP-2608.27513", "title": "DAMP: Decay-Aware Mixed-Precision Recurrent-State Quantization", "version": "v2", "revision": "r1", "tier": "A", "primary_category": "systems", "assessment": "按量化误差及其未来保留程度分配通道精度；改善字节与带宽效率，不直接扩大逻辑地址空间。", "notes": "paper-readings/DAMP-2608.27513/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/DAMP-2608.27513/v2/source/paper.pdf", "source_directory": "paper-readings/DAMP-2608.27513/v2/source", "source_url": "https://arxiv.org/abs/2608.27513v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "DAT", "directory": "DAT-2510.00258", "title": "Delayed Attention Training Improves Length Generalization in Transformer--RNN Hybrids", "version": "v1", "revision": "r1", "tier": "B", "primary_category": "training", "assessment": "先训练循环路径再开放注意力，改善合成复合任务的分工与长度泛化；非大规模 LM 通用验证。", "notes": "paper-readings/DAT-2510.00258/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/DAT-2510.00258/v1/source/paper.pdf", "source_directory": "paper-readings/DAT-2510.00258/v1/source", "source_url": "https://arxiv.org/abs/2510.00258v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "DSLA", "directory": "DSLA-2506.09316", "title": "On-the-Fly Adaptive Distillation of Transformer to Dual-State Linear Attention", "version": "v3", "revision": "r1", "tier": "A", "primary_category": "capacity", "assessment": "历史/近期双状态采用不同遗忘门；容量、时间尺度和逐层转换策略同时改变。", "notes": "paper-readings/DSLA-2506.09316/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/DSLA-2506.09316/v3/source/paper.pdf", "source_directory": "paper-readings/DSLA-2506.09316/v3/source", "source_url": "https://arxiv.org/abs/2506.09316v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "DeltaAttnRes", "directory": "DeltaAttnRes-2605.18855", "title": "Delta Attention Residuals", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/DeltaAttnRes-2605.18855/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/DeltaAttnRes-2605.18855/v1/source/paper.pdf", "source_directory": "paper-readings/DeltaAttnRes-2605.18855/v1/source", "source_url": "https://arxiv.org/abs/2605.18855v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "DeltaLog", "directory": "DeltaLog-2608.15533", "title": "DeltaLog: Deferred Materialization of Recurrent States for Linear Attention Decoding", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "systems", "assessment": "稠密基底加有界更新日志，延迟物理写回；逻辑状态仍逐 token 更新，主要降低写带宽。", "notes": "paper-readings/DeltaLog-2608.15533/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/DeltaLog-2608.15533/v1/source/paper.pdf", "source_directory": "paper-readings/DeltaLog-2608.15533/v1/source", "source_url": "https://arxiv.org/abs/2608.15533v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "DeltaProduct", "directory": "DeltaProduct-2502.10297", "title": "DeltaProduct: Improving State-Tracking in Linear RNNs via Householder Products", "version": "v7", "revision": "r1", "tier": "A", "primary_category": "dynamics", "assessment": "每 token 多次 Householder 型更新；状态维度可不变，表达力来自额外顺序更新与投影。", "notes": "paper-readings/DeltaProduct-2502.10297/v7/r1/reading-notes.zh.md", "pdf": "paper-readings/DeltaProduct-2502.10297/v7/source/paper.pdf", "source_directory": "paper-readings/DeltaProduct-2502.10297/v7/source", "source_url": "https://arxiv.org/abs/2502.10297v7", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "DepthBench", "directory": "DepthBench-2609.32534", "title": "DepthBench: Measuring How Residual Connections Enable More Computational Depth", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/DepthBench-2609.32534/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/DepthBench-2609.32534/v1/source/paper.pdf", "source_directory": "paper-readings/DepthBench-2609.32534/v1/source", "source_url": "https://arxiv.org/abs/2609.32534v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "DepthWeave-KV", "directory": "DepthWeave-KV-2607.06523", "title": "DepthWeave-KV: Token-Adaptive Cross-Layer Residual Factorization for Long-Context KV Cache Compression", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/DepthWeave-KV-2607.06523/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/DepthWeave-KV-2607.06523/v1/source/paper.pdf", "source_directory": "paper-readings/DepthWeave-KV-2607.06523/v1/source", "source_url": "https://arxiv.org/abs/2607.06523v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Diffuser", "directory": "Diffuser-2210.11794", "title": "Diffuser: Efficient Transformers with Multi-hop Attention Diffusion for Long Sequences", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/Diffuser-2210.11794/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/Diffuser-2210.11794/v2/source/paper.pdf", "source_directory": "paper-readings/Diffuser-2210.11794/v2/source", "source_url": "https://arxiv.org/abs/2210.11794v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "DiscoLoop", "directory": "DiscoLoop-2607.00341", "title": "DiscoLoop: Looping Discrete Embeddings and Continuous Hidden States for Multi-hop Reasoning", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/DiscoLoop-2607.00341/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/DiscoLoop-2607.00341/v2/source/paper.pdf", "source_directory": "paper-readings/DiscoLoop-2607.00341/v2/source", "source_url": "https://arxiv.org/abs/2607.00341v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "DoSA", "directory": "DoSA-2610.13990", "title": "DoSA: Dots Sparse Attention", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/DoSA-2610.13990/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/DoSA-2610.13990/v1/source/paper.pdf", "source_directory": "paper-readings/DoSA-2610.13990/v1/source", "source_url": "https://arxiv.org/abs/2610.13990v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "EDA", "directory": "EDA-2606.26560", "title": "Erase-then-Delta Attention: Decoupling Erase and Write Addresses in Delta-Rule Linear Attention", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "geometry", "assessment": "独立擦除地址与写入地址按顺序作用；增加一步清理，不增加同形状内容状态。", "notes": "paper-readings/EDA-2606.26560/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/EDA-2606.26560/v1/source/paper.pdf", "source_directory": "paper-readings/EDA-2606.26560/v1/source", "source_url": "https://arxiv.org/abs/2606.26560v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "EG-MLA", "directory": "EG-MLA-2509.16686", "title": "EG-MLA: Embedding-Gated Multi-head Latent Attention for Scalable and Efficient LLMs", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/EG-MLA-2509.16686/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/EG-MLA-2509.16686/v1/source/paper.pdf", "source_directory": "paper-readings/EG-MLA-2509.16686/v1/source", "source_url": "https://arxiv.org/abs/2509.16686v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "ELF", "directory": "ELF-2605.10938", "title": "ELF: Embedded Language Flows", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/ELF-2605.10938/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/ELF-2605.10938/v2/source/paper.pdf", "source_directory": "paper-readings/ELF-2605.10938/v2/source", "source_url": "https://arxiv.org/abs/2605.10938v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "ELF-REG", "directory": "ELF-REG-2609.29102", "title": "ELF-REG: Scaling Continuous Diffusion Language Models to Reasoning Tasks", "version": "v3", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/ELF-REG-2609.29102/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/ELF-REG-2609.29102/v3/source/paper.pdf", "source_directory": "paper-readings/ELF-REG-2609.29102/v3/source", "source_url": "https://arxiv.org/abs/2609.29102v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "ELR", "directory": "ELR-2610.12582", "title": "One Model to Loop Them All: Elastic Latent Recurrence for Language Models", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/ELR-2610.12582/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/ELR-2610.12582/v1/source/paper.pdf", "source_directory": "paper-readings/ELR-2610.12582/v1/source", "source_url": "https://arxiv.org/abs/2610.12582v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Efficient-DLM", "directory": "Efficient-DLM-2512.14067", "title": "Efficient-DLM: From Autoregressive to Diffusion Language Models, and Beyond in Speed", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/Efficient-DLM-2512.14067/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/Efficient-DLM-2512.14067/v2/source/paper.pdf", "source_directory": "paper-readings/Efficient-DLM-2512.14067/v2/source", "source_url": "https://arxiv.org/abs/2512.14067v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "ExplicitLM", "directory": "ExplicitLM-2511.01581", "title": "ExplicitLM: Decoupling Knowledge from Parameters via Explicit Memory Banks", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "agent-memory", "assessment": "外部文档、智能体、参数记忆、数据或提示组织，未研究本报告的序列内循环 state", "notes": "paper-readings/ExplicitLM-2511.01581/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/ExplicitLM-2511.01581/v1/source/paper.pdf", "source_directory": "paper-readings/ExplicitLM-2511.01581/v1/source", "source_url": "https://arxiv.org/abs/2511.01581v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "FMLM-star", "directory": "FMLM-star-2607.00714", "title": "Self-conditioned Flow Map Language Models via Fixed-point Flows", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/FMLM-star-2607.00714/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/FMLM-star-2607.00714/v2/source/paper.pdf", "source_directory": "paper-readings/FMLM-star-2607.00714/v2/source", "source_url": "https://arxiv.org/abs/2607.00714v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "FRAC", "directory": "FRAC-2609.36314", "title": "Fractional State Space Transition for Long Sequence Modeling", "version": "v1", "revision": "r1", "tier": "B", "primary_category": "dynamics", "assessment": "几何分布的指数时间尺度近似长尾记忆；属于 SSM 近邻，长尾核不等于任意事实无损保留。", "notes": "paper-readings/FRAC-2609.36314/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/FRAC-2609.36314/v1/source/paper.pdf", "source_directory": "paper-readings/FRAC-2609.36314/v1/source", "source_url": "https://arxiv.org/abs/2609.36314v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "FusedKV", "directory": "FusedKV-2512.03870", "title": "Reconstructing KV Caches with Cross-layer Fusion For Enhanced Transformers", "version": "v3", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/FusedKV-2512.03870/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/FusedKV-2512.03870/v3/source/paper.pdf", "source_directory": "paper-readings/FusedKV-2512.03870/v3/source", "source_url": "https://arxiv.org/abs/2512.03870v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "FwPKM", "directory": "FwPKM-2601.00671", "title": "Fast-weight Product Key Memory", "version": "v2", "revision": "r1", "tier": "B", "primary_category": "capacity", "assessment": "乘积键稀疏寻址配合在线快键/快值更新；大容量、重读和真实吞吐成本共同影响结果。", "notes": "paper-readings/FwPKM-2601.00671/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/FwPKM-2601.00671/v2/source/paper.pdf", "source_directory": "paper-readings/FwPKM-2601.00671/v2/source", "source_url": "https://arxiv.org/abs/2601.00671v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "GOAT", "directory": "GOAT-2601.15380", "title": "You Need Better Attention Priors", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/GOAT-2601.15380/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/GOAT-2601.15380/v2/source/paper.pdf", "source_directory": "paper-readings/GOAT-2601.15380/v2/source", "source_url": "https://arxiv.org/abs/2601.15380v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "GenMem", "directory": "GenMem-2609.34633", "title": "GenMem: Generative Symbolic Memory for Self-Evolving Harness", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "agent-memory", "assessment": "外部文档、智能体、参数记忆、数据或提示组织，未研究本报告的序列内循环 state", "notes": "paper-readings/GenMem-2609.34633/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/GenMem-2609.34633/v1/source/paper.pdf", "source_directory": "paper-readings/GenMem-2609.34633/v1/source", "source_url": "https://arxiv.org/abs/2609.34633v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "HOLA", "directory": "HOLA-2607.02303", "title": "A Hippocampus for Linear Attention: An Exact Memory for What the Recurrent State Forgets", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "hybrids", "assessment": "按残差写入幅度选择固定容量精确 KV，并独立归一化读出；兼有新增容量和选择性保护。", "notes": "paper-readings/HOLA-2607.02303/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/HOLA-2607.02303/v1/source/paper.pdf", "source_directory": "paper-readings/HOLA-2607.02303/v1/source", "source_url": "https://arxiv.org/abs/2607.02303v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "HYPIC", "directory": "HYPIC-2607.01299", "title": "HYPIC: Accelerating Hybrid-Attention LLM Serving with Position-Independent Caching", "version": "v2", "revision": "r1", "tier": "A", "primary_category": "systems", "assessment": "缓存段级仿射状态变换以组合历史；固定单层轨迹的精确组合与深层复用近似需区分。", "notes": "paper-readings/HYPIC-2607.01299/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/HYPIC-2607.01299/v2/source/paper.pdf", "source_directory": "paper-readings/HYPIC-2607.01299/v2/source", "source_url": "https://arxiv.org/abs/2607.01299v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "HashAttention", "directory": "HashAttention-2412.14468", "title": "HashAttention: Semantic Sparsity for Faster Inference", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/HashAttention-2412.14468/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/HashAttention-2412.14468/v2/source/paper.pdf", "source_directory": "paper-readings/HashAttention-2412.14468/v2/source", "source_url": "https://arxiv.org/abs/2412.14468v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "HedgeCATs", "directory": "HedgeCATs-2510.05901", "title": "Untangling Component Imbalance in Hybrid Linear Attention Conversion Methods", "version": "v2", "revision": "r1", "tier": "A", "primary_category": "training", "assessment": "揭示滑窗分支掩盖线性分支失效；分阶段迁移及窗口 dropout 促使状态实际参与任务。", "notes": "paper-readings/HedgeCATs-2510.05901/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/HedgeCATs-2510.05901/v2/source/paper.pdf", "source_directory": "paper-readings/HedgeCATs-2510.05901/v2/source", "source_url": "https://arxiv.org/abs/2510.05901v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Hybrid-Hierarchy", "directory": "Hybrid-Hierarchy-2602.01763", "title": "A Provable Expressiveness Hierarchy in Hybrid Linear-Full Attention", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "analysis", "assessment": "有限精度通信下区分状态容量和全注意力访问轮数；专门任务下界不等于一般 LM 的运行成本结论。", "notes": "paper-readings/Hybrid-Hierarchy-2602.01763/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/Hybrid-Hierarchy-2602.01763/v1/source/paper.pdf", "source_directory": "paper-readings/Hybrid-Hierarchy-2602.01763/v1/source", "source_url": "https://arxiv.org/abs/2602.01763v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "HybridPosition", "directory": "HybridPosition-2610.10114", "title": "Mechanics of Long-Context Hybrid Models Part 1.1: From Hybrid Attention to Hybrid Position", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "training", "assessment": "以混合位置偏置解释长上下文模型，SWLA 与 NoPE 缩放改善外推；收益依任务与训练阶段。", "notes": "paper-readings/HybridPosition-2610.10114/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/HybridPosition-2610.10114/v1/source/paper.pdf", "source_directory": "paper-readings/HybridPosition-2610.10114/v1/source", "source_url": "https://arxiv.org/abs/2610.10114v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "HypeNet", "directory": "HypeNet-2601.22156", "title": "Hybrid Linear Attention Done Right: Efficient Distillation and Effective Architectures for Extremely Long Contexts", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "training", "assessment": "少量全局层、循环层和位置编码分工，配合 HALO 蒸馏；是组合设计而非单一状态更新收益。", "notes": "paper-readings/HypeNet-2601.22156/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/HypeNet-2601.22156/v1/source/paper.pdf", "source_directory": "paper-readings/HypeNet-2601.22156/v1/source", "source_url": "https://arxiv.org/abs/2601.22156v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "ICL-Data", "directory": "ICL-Data-2205.05055", "title": "Data Distributional Properties Drive Emergent In-Context Learning in Transformers", "version": "v6", "revision": "r1", "tier": "C", "primary_category": "agent-memory", "assessment": "外部文档、智能体、参数记忆、数据或提示组织，未研究本报告的序列内循环 state", "notes": "paper-readings/ICL-Data-2205.05055/v6/r1/reading-notes.zh.md", "pdf": "paper-readings/ICL-Data-2205.05055/v6/source/paper.pdf", "source_directory": "paper-readings/ICL-Data-2205.05055/v6/source", "source_url": "https://arxiv.org/abs/2205.05055v6", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "JRT", "directory": "JRT-2407.05483", "title": "Just read twice: closing the recall gap for recurrent language models", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "training", "assessment": "问题已知后重读，或训练非因果前缀编码器；从任务条件改变应保存的信息，付出额外输入/计算。", "notes": "paper-readings/JRT-2407.05483/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/JRT-2407.05483/v1/source/paper.pdf", "source_directory": "paper-readings/JRT-2407.05483/v1/source", "source_url": "https://arxiv.org/abs/2407.05483v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "JitMem", "directory": "JitMem-2609.27334", "title": "Just-in-Time Memory: Learning to Curate Task-Adaptive Memory for LLM Agents", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "agent-memory", "assessment": "外部文档、智能体、参数记忆、数据或提示组织，未研究本报告的序列内循环 state", "notes": "paper-readings/JitMem-2609.27334/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/JitMem-2609.27334/v1/source/paper.pdf", "source_directory": "paper-readings/JitMem-2609.27334/v1/source", "source_url": "https://arxiv.org/abs/2609.27334v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "KDN", "directory": "KDN-2609.07816", "title": "Kalman Delta Networks: Uncertainty-aware Associative Memory", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "geometry", "assessment": "显式传播不确定性，以 Kalman gain 决定残差写入；对角/各向同性近似控制辅助状态成本。", "notes": "paper-readings/KDN-2609.07816/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/KDN-2609.07816/v1/source/paper.pdf", "source_directory": "paper-readings/KDN-2609.07816/v1/source", "source_url": "https://arxiv.org/abs/2609.07816v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "KVM", "directory": "KVM-2605.09877", "title": "Key-Value Means: Transformers with Expandable Block-Recurrent Compressed Memory", "version": "v5", "revision": "r1", "tier": "B", "primary_category": "hybrids", "assessment": "滑窗外 KV 按相似度追加/合并压缩槽；固定槽版与随长度增长版具有不同复杂度。", "notes": "paper-readings/KVM-2605.09877/v5/r1/reading-notes.zh.md", "pdf": "paper-readings/KVM-2605.09877/v5/source/paper.pdf", "source_directory": "paper-readings/KVM-2605.09877/v5/source", "source_url": "https://arxiv.org/abs/2605.09877v5", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LAVO", "directory": "LAVO-2312.11135", "title": "Linear Attention via Orthogonal Memory", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "kernels", "assessment": "固定正交方向的均值统计与局部注意力组合；对压缩槽做 softmax，平均压缩会损失历史细节。", "notes": "paper-readings/LAVO-2312.11135/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/LAVO-2312.11135/v1/source/paper.pdf", "source_directory": "paper-readings/LAVO-2312.11135/v1/source", "source_url": "https://arxiv.org/abs/2312.11135v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LLA", "directory": "LLA-2607.15456", "title": "Looped Latent Attention: Cross-Loop KV Compression for Looped Transformers", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/LLA-2607.15456/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/LLA-2607.15456/v2/source/paper.pdf", "source_directory": "paper-readings/LLA-2607.15456/v2/source", "source_url": "https://arxiv.org/abs/2607.15456v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LLM-Memory", "directory": "LLM-Memory-2607.25380", "title": "Memory for Large Language Models", "version": "v1", "revision": "r1", "tier": "B", "primary_category": "analysis", "assessment": "以表示、更新和持续性统一模型级记忆；是综述框架，用来界定状态与外部记忆的边界。", "notes": "paper-readings/LLM-Memory-2607.25380/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/LLM-Memory-2607.25380/v1/source/paper.pdf", "source_directory": "paper-readings/LLM-Memory-2607.25380/v1/source", "source_url": "https://arxiv.org/abs/2607.25380v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LPT", "directory": "LPT-2610.02383", "title": "The Surprising Effectiveness of Shared Memory in Looped Transformers", "version": "v1", "revision": "r2", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/LPT-2610.02383/v1/r2/reading-notes.zh.md", "pdf": "paper-readings/LPT-2610.02383/v1/source/paper.pdf", "source_directory": "paper-readings/LPT-2610.02383/v1/source", "source_url": "https://arxiv.org/abs/2610.02383v1", "available_source_versions": 1, "available_note_revisions": 2 }, { "name": "LRKV", "directory": "LRKV-2601.11471", "title": "Low-Rank Key Value Attention", "version": "v3", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/LRKV-2601.11471/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/LRKV-2601.11471/v3/source/paper.pdf", "source_directory": "paper-readings/LRKV-2601.11471/v3/source", "source_url": "https://arxiv.org/abs/2601.11471v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LaCT", "directory": "LaCT-2505.23884", "title": "Test-Time Training Done Right", "version": "v1", "revision": "r1", "tier": "B", "primary_category": "fastmemory", "assessment": "以大块在线更新提高计算强度，支撑更大非线性快权重；块大小与局部窗口共同影响记忆。", "notes": "paper-readings/LaCT-2505.23884/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/LaCT-2505.23884/v1/source/paper.pdf", "source_directory": "paper-readings/LaCT-2505.23884/v1/source", "source_url": "https://arxiv.org/abs/2505.23884v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LangFlow", "directory": "LangFlow-2604.11748", "title": "LangFlow: Continuous Diffusion Rivals Discrete in Language Modeling", "version": "v3", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/LangFlow-2604.11748/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/LangFlow-2604.11748/v3/source/paper.pdf", "source_directory": "paper-readings/LangFlow-2604.11748/v3/source", "source_url": "https://arxiv.org/abs/2604.11748v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LatentLLM", "directory": "LatentLLM-2505.18413", "title": "LatentLLM: Attention-Aware Joint Tensor Compression", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/LatentLLM-2505.18413/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/LatentLLM-2505.18413/v1/source/paper.pdf", "source_directory": "paper-readings/LatentLLM-2505.18413/v1/source", "source_url": "https://arxiv.org/abs/2505.18413v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Lattice", "directory": "Lattice-2504.05646", "title": "Lattice: Learning to Efficiently Compress the Memory", "version": "v2", "revision": "r1", "tier": "A", "primary_category": "geometry", "assessment": "逐槽正交投影与归一化改进更新几何；局部正交性不等于所有历史关联相互正交。", "notes": "paper-readings/Lattice-2504.05646/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/Lattice-2504.05646/v2/source/paper.pdf", "source_directory": "paper-readings/Lattice-2504.05646/v2/source", "source_url": "https://arxiv.org/abs/2504.05646v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LeapQuant", "directory": "LeapQuant-2609.38166", "title": "LeapQuant: Efficient Linear Attention with Accurate Recurrent State Quantization", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "systems", "assessment": "窗口量化、低秩补偿和残差平滑抑制递归误差；完整预算还应计入高精度日志与补偿因子。", "notes": "paper-readings/LeapQuant-2609.38166/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/LeapQuant-2609.38166/v1/source/paper.pdf", "source_directory": "paper-readings/LeapQuant-2609.38166/v1/source", "source_url": "https://arxiv.org/abs/2609.38166v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LengthGen", "directory": "LengthGen-2507.02782", "title": "Understanding and Improving Length Generalization in Recurrent Models", "version": "v2", "revision": "r2", "tier": "A", "primary_category": "training", "assessment": "状态传递等初态干预扩大训练状态覆盖；无需扩容即可改善外推，稳定 PPL 仍不足以证明远距召回。", "notes": "paper-readings/LengthGen-2507.02782/v2/r2/reading-notes.zh.md", "pdf": "paper-readings/LengthGen-2507.02782/v2/source/paper.pdf", "source_directory": "paper-readings/LengthGen-2507.02782/v2/source", "source_url": "https://arxiv.org/abs/2507.02782v2", "available_source_versions": 1, "available_note_revisions": 2 }, { "name": "LightMem", "directory": "LightMem-2604.07798", "title": "Lightweight LLM Agent Memory with Small Language Models", "version": "v3", "revision": "r1", "tier": "C", "primary_category": "agent-memory", "assessment": "外部文档、智能体、参数记忆、数据或提示组织，未研究本报告的序列内循环 state", "notes": "paper-readings/LightMem-2604.07798/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/LightMem-2604.07798/v3/source/paper.pdf", "source_directory": "paper-readings/LightMem-2604.07798/v3/source", "source_url": "https://arxiv.org/abs/2604.07798v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LinearAttn-Memory", "directory": "LinearAttn-Memory-2609.33093", "title": "How Linear Attention Remembers", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "analysis", "assessment": "因果干预区分写入、保留、读出；负载与写入重叠损害召回，混合模型直接记忆载体可转向 KV。", "notes": "paper-readings/LinearAttn-Memory-2609.33093/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/LinearAttn-Memory-2609.33093/v1/source/paper.pdf", "source_directory": "paper-readings/LinearAttn-Memory-2609.33093/v1/source", "source_url": "https://arxiv.org/abs/2609.33093v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LoLA", "directory": "LoLA-2505.23666", "title": "LoLA: Low-Rank Linear Attention With Sparse Caching", "version": "v3", "revision": "r1", "tier": "A", "primary_category": "hybrids", "assessment": "按自召回误差将难压缩 KV 放入有限全局缓存；等名义缓存预算对照支持分配策略的价值。", "notes": "paper-readings/LoLA-2505.23666/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/LoLA-2505.23666/v3/source/paper.pdf", "source_directory": "paper-readings/LoLA-2505.23666/v3/source", "source_url": "https://arxiv.org/abs/2505.23666v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LoLCATs", "directory": "LoLCATs-2410.10254", "title": "LoLCATs: On Low-Rank Linearizing of Large Language Models", "version": "v3", "revision": "r1", "tier": "A", "primary_category": "training", "assessment": "先迁移注意力输出、再用 LoRA 适配混合模型；转换成本降低，总体分数需配合分支消融解读。", "notes": "paper-readings/LoLCATs-2410.10254/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/LoLCATs-2410.10254/v3/source/paper.pdf", "source_directory": "paper-readings/LoLCATs-2410.10254/v3/source", "source_url": "https://arxiv.org/abs/2410.10254v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LoopLM-Conditioning", "directory": "LoopLM-Conditioning-2609.36636", "title": "What Makes Recurrence Effective in Looped Language Models?", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/LoopLM-Conditioning-2609.36636/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/LoopLM-Conditioning-2609.36636/v1/source/paper.pdf", "source_directory": "paper-readings/LoopLM-Conditioning-2609.36636/v1/source", "source_url": "https://arxiv.org/abs/2609.36636v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LoopMDM", "directory": "LoopMDM-2605.26106", "title": "Looped Diffusion Language Models", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/LoopMDM-2605.26106/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/LoopMDM-2605.26106/v1/source/paper.pdf", "source_directory": "paper-readings/LoopMDM-2605.26106/v1/source", "source_url": "https://arxiv.org/abs/2605.26106v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "LoopedReuse", "directory": "LoopedReuse-2610.19768", "title": "Learning to Compute Is Not Learning to Retain or Reuse in Looped Transformers", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/LoopedReuse-2610.19768/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/LoopedReuse-2610.19768/v1/source/paper.pdf", "source_directory": "paper-readings/LoopedReuse-2610.19768/v1/source", "source_url": "https://arxiv.org/abs/2610.19768v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "M2RNN", "directory": "M2RNN-2603.14360", "title": "M$^2$RNN: Non-Linear RNNs with Matrix-Valued States for Scalable Language Modeling", "version": "v2", "revision": "r1", "tier": "B", "primary_category": "dynamics", "assessment": "矩阵状态加入逐步非线性递推；改善少量替换层的混合模型，牺牲标准线性扫描结构。", "notes": "paper-readings/M2RNN-2603.14360/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/M2RNN-2603.14360/v2/source/paper.pdf", "source_directory": "paper-readings/M2RNN-2603.14360/v2/source", "source_url": "https://arxiv.org/abs/2603.14360v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "MDM-LS", "directory": "MDM-LS-2511.19152", "title": "Masked Diffusion Models are Secretly Learned-Order Autoregressive Models", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/MDM-LS-2511.19152/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/MDM-LS-2511.19152/v1/source/paper.pdf", "source_directory": "paper-readings/MDM-LS-2511.19152/v1/source", "source_url": "https://arxiv.org/abs/2511.19152v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "MDN", "directory": "MDN-2605.05838", "title": "MDN: Parallelizing Stepwise Momentum for Delta Linear Attention", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "capacity", "assessment": "用额外动量矩阵积累历史残差，提供逐步动量的分块算法；同尺寸下主状态加动量约为双状态。", "notes": "paper-readings/MDN-2605.05838/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/MDN-2605.05838/v1/source/paper.pdf", "source_directory": "paper-readings/MDN-2605.05838/v1/source", "source_url": "https://arxiv.org/abs/2605.05838v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "MELT", "directory": "MELT-2605.07721", "title": "Memory-Efficient Looped Transformer: Decoupling Compute from Memory in Looped Language Models", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/MELT-2605.07721/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/MELT-2605.07721/v2/source/paper.pdf", "source_directory": "paper-readings/MELT-2605.07721/v2/source", "source_url": "https://arxiv.org/abs/2605.07721v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "MS-GLA", "directory": "MS-GLA-2609.35664", "title": "MS-GLA: Multi-Scale Gated Linear Attention for Addressing Representational Bottlenecks via Multi-Temporal Resolution", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "dynamics", "assessment": "固定总头数分配给不同池化尺度，延迟后因果融合；用时间分辨率换取低频写入和保护。", "notes": "paper-readings/MS-GLA-2609.35664/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/MS-GLA-2609.35664/v1/source/paper.pdf", "source_directory": "paper-readings/MS-GLA-2609.35664/v1/source", "source_url": "https://arxiv.org/abs/2609.35664v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Mamba-3", "directory": "Mamba-3-2603.15569", "title": "Mamba-3: Improved Sequence Modeling using State Space Principles", "version": "v1", "revision": "r1", "tier": "B", "primary_category": "dynamics", "assessment": "指数梯形写入、旋转与 MIMO 组合；MIMO 可在同状态上增加读写计算，训练 FLOPs 随之增加。", "notes": "paper-readings/Mamba-3-2603.15569/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/Mamba-3-2603.15569/v1/source/paper.pdf", "source_directory": "paper-readings/Mamba-3-2603.15569/v1/source", "source_url": "https://arxiv.org/abs/2603.15569v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Mamba3-Oscillation", "directory": "Mamba3-Oscillation-2610.13341", "title": "Understanding Mamba-3 Through Oscillatory Dynamics", "version": "v1", "revision": "r1", "tier": "B", "primary_category": "dynamics", "assessment": "以振荡动力学和相位干预解释旋转如何影响累积与读出；机制分析，不是另一套扩容模型。", "notes": "paper-readings/Mamba3-Oscillation-2610.13341/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/Mamba3-Oscillation-2610.13341/v1/source/paper.pdf", "source_directory": "paper-readings/Mamba3-Oscillation-2610.13341/v1/source", "source_url": "https://arxiv.org/abs/2610.13341v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "MemPrism", "directory": "MemPrism-2608.06745", "title": "MemPrism: Task-Conditioned Relational Memory Views for Long-Horizon Agents", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "agent-memory", "assessment": "外部文档、智能体、参数记忆、数据或提示组织，未研究本报告的序列内循环 state", "notes": "paper-readings/MemPrism-2608.06745/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/MemPrism-2608.06745/v1/source/paper.pdf", "source_directory": "paper-readings/MemPrism-2608.06745/v1/source", "source_url": "https://arxiv.org/abs/2608.06745v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "MemoryMosaics", "directory": "MemoryMosaics-2507.03285", "title": "Memory Mosaics at scale", "version": "v3", "revision": "r1", "tier": "B", "primary_category": "fastmemory", "assessment": "自适应核、门控 key 和多层记忆改善 ICL；长期分支保存历史 KV，不属于固定状态线性注意力。", "notes": "paper-readings/MemoryMosaics-2507.03285/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/MemoryMosaics-2507.03285/v3/source/paper.pdf", "source_directory": "paper-readings/MemoryMosaics-2507.03285/v3/source", "source_url": "https://arxiv.org/abs/2507.03285v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "MesaNet", "directory": "MesaNet-2506.05233", "title": "MesaNet: Sequence Modeling by Locally Optimal Test-Time Training", "version": "v2", "revision": "r1", "tier": "A", "primary_category": "geometry", "assessment": "累计交叉统计与 Gram 统计，读取时用 CG 求回归解；更好局部拟合需要辅助矩阵与求解计算。", "notes": "paper-readings/MesaNet-2506.05233/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/MesaNet-2506.05233/v2/source/paper.pdf", "source_directory": "paper-readings/MesaNet-2506.05233/v2/source", "source_url": "https://arxiv.org/abs/2506.05233v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Metis", "directory": "Metis-2607.26760", "title": "Metis: Memory Foundation Model", "version": "v2", "revision": "r1", "tier": "B", "primary_category": "fastmemory", "assessment": "跨轮持久矩阵以学习到的前向规则读写；说明紧凑状态可用于持久记忆，但离完整上下文质量仍远。", "notes": "paper-readings/Metis-2607.26760/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/Metis-2607.26760/v2/source/paper.pdf", "source_directory": "paper-readings/Metis-2607.26760/v2/source", "source_url": "https://arxiv.org/abs/2607.26760v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "MiTA", "directory": "MiTA-2602.01219", "title": "Mixture-of-Top-k Attention: Efficient Attention via Scalable Fast Weights", "version": "v5", "revision": "r1", "tier": "B", "primary_category": "hybrids", "assessment": "landmark 复用 top-k KV 子集并加入全局压缩记忆；主要验证于视觉与 LRA，非固定状态 LM 直接对照。", "notes": "paper-readings/MiTA-2602.01219/v5/r1/reading-notes.zh.md", "pdf": "paper-readings/MiTA-2602.01219/v5/source/paper.pdf", "source_directory": "paper-readings/MiTA-2602.01219/v5/source", "source_url": "https://arxiv.org/abs/2602.01219v5", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "MoD", "directory": "MoD-2404.02258", "title": "Mixture-of-Depths: Dynamically allocating compute in transformer-based language models", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/MoD-2404.02258/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/MoD-2404.02258/v1/source/paper.pdf", "source_directory": "paper-readings/MoD-2404.02258/v1/source", "source_url": "https://arxiv.org/abs/2404.02258v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "MoDA", "directory": "MoDA-2603.15619", "title": "Mixture-of-Depths Attention", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/MoDA-2603.15619/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/MoDA-2603.15619/v1/source/paper.pdf", "source_directory": "paper-readings/MoDA-2603.15619/v1/source", "source_url": "https://arxiv.org/abs/2603.15619v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "MoM", "directory": "MoM-2502.13685", "title": "MoM: Linear Sequence Modeling with Mixture-of-Memories", "version": "v4", "revision": "r1", "tier": "A", "primary_category": "capacity", "assessment": "多个独立循环记忆、top-k 路由及共享记忆；同时扩容与隔离写入，等激活预算不代表等总状态。", "notes": "paper-readings/MoM-2502.13685/v4/r1/reading-notes.zh.md", "pdf": "paper-readings/MoM-2502.13685/v4/source/paper.pdf", "source_directory": "paper-readings/MoM-2502.13685/v4/source", "source_url": "https://arxiv.org/abs/2502.13685v4", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "MoNe", "directory": "MoNe-2608.17616", "title": "MoNe: Modular Neural Memory for Efficient Long Context Inference", "version": "v1", "revision": "r1", "tier": "B", "primary_category": "fastmemory", "assessment": "冻结 Transformer，逐层 SwiGLU 快权重压缩上下文，再转成查询 KV；任务训练与受控检索是重要条件。", "notes": "paper-readings/MoNe-2608.17616/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/MoNe-2608.17616/v1/source/paper.pdf", "source_directory": "paper-readings/MoNe-2608.17616/v1/source", "source_url": "https://arxiv.org/abs/2608.17616v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "NAtS-L", "directory": "NAtS-L-2602.03681", "title": "Neural Attention Search Linear: Towards Adaptive Token-Level Hybrid Attention Models", "version": "v2", "revision": "r1", "tier": "A", "primary_category": "hybrids", "assessment": "按 64-token 块选择显式 KV 或 GDN，并进行 query 条件融合；无硬 KV 预算时仍含二次项。", "notes": "paper-readings/NAtS-L-2602.03681/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/NAtS-L-2602.03681/v2/source/paper.pdf", "source_directory": "paper-readings/NAtS-L-2602.03681/v2/source", "source_url": "https://arxiv.org/abs/2602.03681v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "NCP-ArchPreview", "directory": "NCP-ArchPreview-2609.10715", "title": "NCP-ArchPreview Technical Report: Moving towards Latent Space Language Models through Next Concept Prediction", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/NCP-ArchPreview-2609.10715/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/NCP-ArchPreview-2609.10715/v1/source/paper.pdf", "source_directory": "paper-readings/NCP-ArchPreview-2609.10715/v1/source", "source_url": "https://arxiv.org/abs/2609.10715v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "NL", "directory": "NL-2512.24695", "title": "Nested Learning: The Illusion of Deep Learning Architectures", "version": "v1", "revision": "r1", "tier": "B", "primary_category": "fastmemory", "assessment": "多时间尺度优化统一记忆和优化器，Hope 组合自指记忆与 CMS；容量与优化器状态均应计入。", "notes": "paper-readings/NL-2512.24695/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/NL-2512.24695/v1/source/paper.pdf", "source_directory": "paper-readings/NL-2512.24695/v1/source", "source_url": "https://arxiv.org/abs/2512.24695v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Nectar", "directory": "Nectar-2605.09778", "title": "Nectar: Neural Estimation of Cached-Token Attention via Regression", "version": "v2", "revision": "r1", "tier": "B", "primary_category": "hybrids", "assessment": "离线用小网络拟合固定文档的注意力输出与归一化量；依赖查询分布，非在线流式更新规则。", "notes": "paper-readings/Nectar-2605.09778/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/Nectar-2605.09778/v2/source/paper.pdf", "source_directory": "paper-readings/Nectar-2605.09778/v2/source", "source_url": "https://arxiv.org/abs/2605.09778v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "OAL", "directory": "OAL-2610.13674", "title": "Not All Dimensions Are Equal: Outlier-Aware Attention Linearization with Grouped Quadratic Kernels", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "training", "assessment": "按点积贡献分组学习二次核；压缩可学习系数，当前实现仍保存原始维度对的二阶状态。", "notes": "paper-readings/OAL-2610.13674/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/OAL-2610.13674/v1/source/paper.pdf", "source_directory": "paper-readings/OAL-2610.13674/v1/source", "source_url": "https://arxiv.org/abs/2610.13674v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "OASIS", "directory": "OASIS-2605.17887", "title": "Attention Sinks and Outliers in Attention Residuals", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/OASIS-2605.17887/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/OASIS-2605.17887/v2/source/paper.pdf", "source_directory": "paper-readings/OASIS-2605.17887/v2/source", "source_url": "https://arxiv.org/abs/2605.17887v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "OSDN", "directory": "OSDN-2605.13473", "title": "OSDN: Improving Delta Rule with Provable Online Preconditioning in Linear Attention", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "geometry", "assessment": "每头增加一个对角预条件向量，区分残差读取键与写回键；局部残差改善未必带来召回提升。", "notes": "paper-readings/OSDN-2605.13473/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/OSDN-2605.13473/v1/source/paper.pdf", "source_directory": "paper-readings/OSDN-2605.13473/v1/source", "source_url": "https://arxiv.org/abs/2605.13473v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "PDN", "directory": "PDN-2604.21100", "title": "Preconditioned DeltaNet: Curvature-aware Sequence Modeling for Linear Recurrences", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "geometry", "assessment": "历史 key 的对角统计调整写 key；精确逆 Gram 的回归等价不能直接套到实用对角近似。", "notes": "paper-readings/PDN-2604.21100/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/PDN-2604.21100/v1/source/paper.pdf", "source_directory": "paper-readings/PDN-2604.21100/v1/source", "source_url": "https://arxiv.org/abs/2604.21100v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "PaTH", "directory": "PaTH-2505.16381", "title": "PaTH Attention: Position Encoding via Accumulating Householder Transformations", "version": "v2", "revision": "r1", "tier": "B", "primary_category": "dynamics", "assessment": "把 Householder 路径转移用于 softmax 的位置匹配；保留显式历史和二次项，是状态追踪参照。", "notes": "paper-readings/PaTH-2505.16381/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/PaTH-2505.16381/v2/source/paper.pdf", "source_directory": "paper-readings/PaTH-2505.16381/v2/source", "source_url": "https://arxiv.org/abs/2505.16381v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Parcae", "directory": "Parcae-2604.12946", "title": "Parcae: Scaling Laws For Stable Looped Language Models", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/Parcae-2604.12946/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/Parcae-2604.12946/v1/source/paper.pdf", "source_directory": "paper-readings/Parcae-2604.12946/v1/source", "source_url": "https://arxiv.org/abs/2604.12946v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "PartRep", "directory": "PartRep-2607.01792", "title": "PARTREP: Learning What to Repeat for Decoder-only LLMs", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "agent-memory", "assessment": "外部文档、智能体、参数记忆、数据或提示组织，未研究本报告的序列内循环 state", "notes": "paper-readings/PartRep-2607.01792/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/PartRep-2607.01792/v1/source/paper.pdf", "source_directory": "paper-readings/PartRep-2607.01792/v1/source", "source_url": "https://arxiv.org/abs/2607.01792v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Proteus", "directory": "Proteus-2608.16844", "title": "Proteus: Incremental Memory Activation for Long-Context Sequence Modeling", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "geometry", "assessment": "在固定最大记忆内逐步开放读写块；改变容量使用时机，超过训练窗口后容量不会继续增长。", "notes": "paper-readings/Proteus-2608.16844/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/Proteus-2608.16844/v1/source/paper.pdf", "source_directory": "paper-readings/Proteus-2608.16844/v1/source", "source_url": "https://arxiv.org/abs/2608.16844v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Puro-2B", "directory": "Puro-2B-2608.27370", "title": "Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/Puro-2B-2608.27370/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/Puro-2B-2608.27370/v2/source/paper.pdf", "source_directory": "paper-readings/Puro-2B-2608.27370/v2/source", "source_url": "https://arxiv.org/abs/2608.27370v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Q-Delta", "directory": "Q-Delta-2606.08804", "title": "Q-Delta: Beyond Key-Value Associative State Evolution", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "geometry", "assessment": "将 query 的旧状态读出反馈到写入残差；内容状态形状保持，更新表达力和计算增加。", "notes": "paper-readings/Q-Delta-2606.08804/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/Q-Delta-2606.08804/v1/source/paper.pdf", "source_directory": "paper-readings/Q-Delta-2606.08804/v1/source", "source_url": "https://arxiv.org/abs/2606.08804v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "RACE", "directory": "RACE-2510.04008", "title": "RACE Attention: A Strictly Linear-Time Attention Layer for Training on Outrageously Large Contexts", "version": "v5", "revision": "r1", "tier": "A", "primary_category": "kernels", "assessment": "可微软哈希桶保存质量与值统计；固定桶数下线性，容量、碰撞和哈希构造仍有成本。", "notes": "paper-readings/RACE-2510.04008/v5/r1/reading-notes.zh.md", "pdf": "paper-readings/RACE-2510.04008/v5/source/paper.pdf", "source_directory": "paper-readings/RACE-2510.04008/v5/source", "source_url": "https://arxiv.org/abs/2510.04008v5", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "REMORY", "directory": "REMORY-2610.11287", "title": "REMORY: Learning Residual Memory for Context Compaction", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "agent-memory", "assessment": "外部文档、智能体、参数记忆、数据或提示组织，未研究本报告的序列内循环 state", "notes": "paper-readings/REMORY-2610.11287/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/REMORY-2610.11287/v1/source/paper.pdf", "source_directory": "paper-readings/REMORY-2610.11287/v1/source", "source_url": "https://arxiv.org/abs/2610.11287v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "RIS-Kernel", "directory": "RIS-Kernel-2607.21927", "title": "RIS-Kernel: A Model-Agnostic Architecture for Long-Context LLM Inference via Sparse Attention", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/RIS-Kernel-2607.21927/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/RIS-Kernel-2607.21927/v1/source/paper.pdf", "source_directory": "paper-readings/RIS-Kernel-2607.21927/v1/source", "source_url": "https://arxiv.org/abs/2607.21927v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "RelaxedRecursive", "directory": "RelaxedRecursive-2410.20672", "title": "Relaxed Recursive Transformers: Effective Parameter Sharing with Layer-wise LoRA", "version": "v3", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/RelaxedRecursive-2410.20672/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/RelaxedRecursive-2410.20672/v3/source/paper.pdf", "source_directory": "paper-readings/RelaxedRecursive-2410.20672/v3/source", "source_url": "https://arxiv.org/abs/2410.20672v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "ResFormer", "directory": "ResFormer-2410.17897", "title": "Value Residual Learning", "version": "v5", "revision": "r1", "tier": "C", "primary_category": "depth-recurrence", "assessment": "网络深度、层间/循环间共享及投影结构，未作为 token 历史 state 方法计数", "notes": "paper-readings/ResFormer-2410.17897/v5/r1/reading-notes.zh.md", "pdf": "paper-readings/ResFormer-2410.17897/v5/source/paper.pdf", "source_directory": "paper-readings/ResFormer-2410.17897/v5/source", "source_url": "https://arxiv.org/abs/2410.17897v5", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "ResidualKV", "directory": "ResidualKV-2602.08005", "title": "ResidualKV: Residual-Based KV Cache Compression for Efficient Long-Context Inference", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/ResidualKV-2602.08005/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/ResidualKV-2602.08005/v2/source/paper.pdf", "source_directory": "paper-readings/ResidualKV-2602.08005/v2/source", "source_url": "https://arxiv.org/abs/2602.08005v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "SANATAN", "directory": "SANATAN-github", "title": "SANATAN: Root-Parity Gated Angular Attention with an Exact Constant-State Recurrent Form", "version": "v2", "revision": "r1", "tier": "A", "primary_category": "kernels", "assessment": "角度核及根标签门控具有显式 6d 特征；常状态精确递推针对 p=1，强基线和公开制品尚有差距。", "notes": "paper-readings/SANATAN-github/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/SANATAN-github/v2/source/paper.pdf", "source_directory": "paper-readings/SANATAN-github/v2/source", "source_url": "https://github.com/bhargavpatel431997/sanatan-attention", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "SAS", "directory": "SAS-2609.13141", "title": "SAS: Simple Attention Sparsification via End-to-End Optimization of Context Ranking", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/SAS-2609.13141/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/SAS-2609.13141/v1/source/paper.pdf", "source_directory": "paper-readings/SAS-2609.13141/v1/source", "source_url": "https://arxiv.org/abs/2609.13141v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "SDM", "directory": "SDM-2607.07386", "title": "Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "capacity", "assessment": "大槽表配合稀疏 top-k 读写；容量和激活算术部分解耦，但可学习初态、总存储及路由成本须单列。", "notes": "paper-readings/SDM-2607.07386/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/SDM-2607.07386/v1/source/paper.pdf", "source_directory": "paper-readings/SDM-2607.07386/v1/source", "source_url": "https://arxiv.org/abs/2607.07386v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "SEDD", "directory": "SEDD-2310.16834", "title": "Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution", "version": "v3", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/SEDD-2310.16834/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/SEDD-2310.16834/v3/source/paper.pdf", "source_directory": "paper-readings/SEDD-2310.16834/v3/source", "source_url": "https://arxiv.org/abs/2310.16834v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "SFDA", "directory": "SFDA-2607.11897", "title": "Semidirect Fourier Delta Attention: Phase-Controlled Delta Memory with Constructive Chunk-WY Kernels", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "dynamics", "assessment": "相位旋转扩展 delta 记忆并构造精确块 WY；证据主要为理论和小型计数实验，非大规模 LM 验证。", "notes": "paper-readings/SFDA-2607.11897/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/SFDA-2607.11897/v1/source/paper.pdf", "source_directory": "paper-readings/SFDA-2607.11897/v1/source", "source_url": "https://arxiv.org/abs/2607.11897v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "SRWM", "directory": "SRWM-2202.05780", "title": "A Modern Self-Referential Weight Matrix That Learns to Modify Itself", "version": "v2", "revision": "r1", "tier": "B", "primary_category": "fastmemory", "assessment": "同一矩阵产生并执行自身 delta 更新；扩展的是更新控制器的适应能力，而非单纯槽位数。", "notes": "paper-readings/SRWM-2202.05780/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/SRWM-2202.05780/v2/source/paper.pdf", "source_directory": "paper-readings/SRWM-2202.05780/v2/source", "source_url": "https://arxiv.org/abs/2202.05780v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "SSM-Gating", "directory": "SSM-Gating-2609.16540", "title": "On the Importance of Gating: Memorization vs. In-Context Learning in State Space Models", "version": "v1", "revision": "r1", "tier": "B", "primary_category": "analysis", "assessment": "区分门控导致的优化捷径与长历史干扰；遗忘既可能削弱信号，也可能抑制噪声。", "notes": "paper-readings/SSM-Gating-2609.16540/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/SSM-Gating-2609.16540/v1/source/paper.pdf", "source_directory": "paper-readings/SSM-Gating-2609.16540/v1/source", "source_url": "https://arxiv.org/abs/2609.16540v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Samba", "directory": "Samba-2406.07522", "title": "Samba: Simple Hybrid State Space Models for Efficient Unlimited Context Language Modeling", "version": "v3", "revision": "r1", "tier": "B", "primary_category": "hybrids", "assessment": "Mamba 与滑窗注意力交替；窗口承担精确局部检索，长文 PPL 外推与精确召回分开评估。", "notes": "paper-readings/Samba-2406.07522/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/Samba-2406.07522/v3/source/paper.pdf", "source_directory": "paper-readings/Samba-2406.07522/v3/source", "source_url": "https://arxiv.org/abs/2406.07522v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "SelectiveRoPE", "directory": "SelectiveRoPE-2511.17388", "title": "Selective Rotary Position Embedding", "version": "v3", "revision": "r1", "tier": "A", "primary_category": "dynamics", "assessment": "输入相关累计旋转改变读写匹配；不靠增加槽数，语言模型收益随架构与规模变化。", "notes": "paper-readings/SelectiveRoPE-2511.17388/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/SelectiveRoPE-2511.17388/v3/source/paper.pdf", "source_directory": "paper-readings/SelectiveRoPE-2511.17388/v3/source", "source_url": "https://arxiv.org/abs/2511.17388v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Sigma", "directory": "Sigma-2610.02665", "title": "Large Language Continuous Diffusion Models", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/Sigma-2610.02665/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/Sigma-2610.02665/v1/source/paper.pdf", "source_directory": "paper-readings/Sigma-2610.02665/v1/source", "source_url": "https://arxiv.org/abs/2610.02665v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Sleep", "directory": "Sleep-2606.03979", "title": "Language Models Need Sleep: Learning to Self-Modify and Consolidate Memories", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "agent-memory", "assessment": "外部文档、智能体、参数记忆、数据或提示组织，未研究本报告的序列内循环 state", "notes": "paper-readings/Sleep-2606.03979/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/Sleep-2606.03979/v2/source/paper.pdf", "source_directory": "paper-readings/Sleep-2606.03979/v2/source", "source_url": "https://arxiv.org/abs/2606.03979v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Smoothie", "directory": "Smoothie-2505.18853", "title": "Smoothie: Smoothing Diffusion on Token Embeddings for Text Generation", "version": "v2", "revision": "r1", "tier": "C", "primary_category": "other", "assessment": "主要研究扩散、潜空间语言建模、训练或表示分析，未给出本文所需的 state 机制", "notes": "paper-readings/Smoothie-2505.18853/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/Smoothie-2505.18853/v2/source/paper.pdf", "source_directory": "paper-readings/Smoothie-2505.18853/v2/source", "source_url": "https://arxiv.org/abs/2505.18853v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "SparDA", "directory": "SparDA-2606.04511", "title": "SparDA: Sparse Decoupled Attention for Efficient Long-Context LLM Inference", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/SparDA-2606.04511/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/SparDA-2606.04511/v1/source/paper.pdf", "source_directory": "paper-readings/SparDA-2606.04511/v1/source", "source_url": "https://arxiv.org/abs/2606.04511v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "SpargeAttn", "directory": "SpargeAttn-2502.18137", "title": "SpargeAttention: Accurate and Training-free Sparse Attention Accelerating Any Model Inference", "version": "v8", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/SpargeAttn-2502.18137/v8/r1/reading-notes.zh.md", "pdf": "paper-readings/SpargeAttn-2502.18137/v8/source/paper.pdf", "source_directory": "paper-readings/SpargeAttn-2502.18137/v8/source", "source_url": "https://arxiv.org/abs/2502.18137v8", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "SpectralShift", "directory": "SpectralShift-2609.14320", "title": "SpectralShift: Effective Context Window Extension of Gated DeltaNet via Spectral Reparameterization", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "dynamics", "assessment": "扩窗时重参数化遗忘门并降低相应学习率，改善慢衰减方向；无新增内容状态。", "notes": "paper-readings/SpectralShift-2609.14320/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/SpectralShift-2609.14320/v1/source/paper.pdf", "source_directory": "paper-readings/SpectralShift-2609.14320/v1/source", "source_url": "https://arxiv.org/abs/2609.14320v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "StateReduction", "directory": "StateReduction-2602.04852", "title": "On State Reduction in Linear Attention", "version": "v3", "revision": "r1", "tier": "A", "primary_category": "systems", "assessment": "联合裁剪 key/query 通道以缩小真实状态；恢复后 PPL 接近原模型，纯递归长距召回仍可能下降。", "notes": "paper-readings/StateReduction-2602.04852/v3/r1/reading-notes.zh.md", "pdf": "paper-readings/StateReduction-2602.04852/v3/source/paper.pdf", "source_directory": "paper-readings/StateReduction-2602.04852/v3/source", "source_url": "https://arxiv.org/abs/2602.04852v3", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "TIDE", "directory": "TIDE-2605.06216", "title": "TIDE: Every Layer Knows the Token Beneath the Context", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "agent-memory", "assessment": "外部文档、智能体、参数记忆、数据或提示组织，未研究本报告的序列内循环 state", "notes": "paper-readings/TIDE-2605.06216/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/TIDE-2605.06216/v1/source/paper.pdf", "source_directory": "paper-readings/TIDE-2605.06216/v1/source", "source_url": "https://arxiv.org/abs/2605.06216v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "TTT", "directory": "TTT-2407.04620", "title": "Learning to (Learn at Test Time): RNNs with Expressive Hidden States", "version": "v4", "revision": "r1", "tier": "B", "primary_category": "fastmemory", "assessment": "将小模型权重作为隐状态，由自监督内层目标更新；线性/MLP 状态的表达力、计算与存储不同。", "notes": "paper-readings/TTT-2407.04620/v4/r1/reading-notes.zh.md", "pdf": "paper-readings/TTT-2407.04620/v4/source/paper.pdf", "source_directory": "paper-readings/TTT-2407.04620/v4/source", "source_url": "https://arxiv.org/abs/2407.04620v4", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Tail-Replay", "directory": "Tail-Replay-2608.30310", "title": "Tail-Replay: Escaping the Curse of Linear Attention in Prefix Caching for Hybrid LLMs", "version": "v2", "revision": "r1", "tier": "A", "primary_category": "systems", "assessment": "缓存精确 FA 输出并短尾重放恢复 LA 状态；改善前缀复用，增加辅助缓存且恢复存在误差。", "notes": "paper-readings/Tail-Replay-2608.30310/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/Tail-Replay-2608.30310/v2/source/paper.pdf", "source_directory": "paper-readings/Tail-Replay-2608.30310/v2/source", "source_url": "https://arxiv.org/abs/2608.30310v2", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Trace-as-State", "directory": "Trace-as-State-2609.02702", "title": "Trace as State: Reasoning Traces as Conditional States for Long-Context Transformers", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "agent-memory", "assessment": "外部文档、智能体、参数记忆、数据或提示组织，未研究本报告的序列内循环 state", "notes": "paper-readings/Trace-as-State-2609.02702/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/Trace-as-State-2609.02702/v1/source/paper.pdf", "source_directory": "paper-readings/Trace-as-State-2609.02702/v1/source", "source_url": "https://arxiv.org/abs/2609.02702v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Transformer-Topology", "directory": "Transformer-Topology-2604.17121", "title": "The Topological Trouble With Transformers", "version": "v5", "revision": "r1", "tier": "B", "primary_category": "analysis", "assessment": "分析计算图中状态跨位置/跨层的可达性；为增加状态计算和循环深度提供结构参照。", "notes": "paper-readings/Transformer-Topology-2604.17121/v5/r1/reading-notes.zh.md", "pdf": "paper-readings/Transformer-Topology-2604.17121/v5/source/paper.pdf", "source_directory": "paper-readings/Transformer-Topology-2604.17121/v5/source", "source_url": "https://arxiv.org/abs/2604.17121v5", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "Triadic", "directory": "Triadic-2609.36529", "title": "Triadic Linear Attention: Three-Dimensional Recurrent States for Long-Context Sequence Modeling", "version": "v1", "revision": "r1", "tier": "A", "primary_category": "capacity", "assessment": "双 key 的张量积扩展联合地址，状态增至 E 倍；少量新增投影不消除稠密状态的存储和读写成本。", "notes": "paper-readings/Triadic-2609.36529/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/Triadic-2609.36529/v1/source/paper.pdf", "source_directory": "paper-readings/Triadic-2609.36529/v1/source", "source_url": "https://arxiv.org/abs/2609.36529v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "VHA", "directory": "VHA-2610.17959", "title": "Virtual Head Attention", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/VHA-2610.17959/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/VHA-2610.17959/v1/source/paper.pdf", "source_directory": "paper-readings/VHA-2610.17959/v1/source", "source_url": "https://arxiv.org/abs/2610.17959v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "VarRate", "directory": "VarRate-2607.15498", "title": "VarRate: Training-Free Variable-Rate KV Cache Compression for Long-Context LLMs", "version": "v1", "revision": "r1", "tier": "C", "primary_category": "attention-comparison", "assessment": "显式 KV、稀疏检索与压缩的参照，未作为固定循环 state 方法计数", "notes": "paper-readings/VarRate-2607.15498/v1/r1/reading-notes.zh.md", "pdf": "paper-readings/VarRate-2607.15498/v1/source/paper.pdf", "source_directory": "paper-readings/VarRate-2607.15498/v1/source", "source_url": "https://arxiv.org/abs/2607.15498v1", "available_source_versions": 1, "available_note_revisions": 1 }, { "name": "YOCO", "directory": "YOCO-2405.05254", "title": "You Only Cache Once: Decoder-Decoder Architectures for Language Models", "version": "v2", "revision": "r1", "tier": "B", "primary_category": "hybrids", "assessment": "高效 self-decoder 生成一份跨层共享的全局 KV；包含循环模块，但全模型状态仍随历史增长。", "notes": "paper-readings/YOCO-2405.05254/v2/r1/reading-notes.zh.md", "pdf": "paper-readings/YOCO-2405.05254/v2/source/paper.pdf", "source_directory": "paper-readings/YOCO-2405.05254/v2/source", "source_url": "https://arxiv.org/abs/2405.05254v2", "available_source_versions": 1, "available_note_revisions": 1 } ] }