📝 Drafts

论文学习草稿列表 · 与 42 协作中暂存的内容

轨迹压缩

4 篇
📅 2026-05-20 📖 citations 🔗 GitHub 🔖 ID: 2605.23296
💡 96k token 输入→~3× 输出(vs 线性增长),块级别细粒度控制
parallel compaction block-based compaction volume control long-horizon serving
提供细粒度 summary volume 控制,与 sequential baseline 对比
📅 2026-05-10 📖 citations 🔗 GitHub 🔖 ID: 2605.0858
💡 异步 compaction 解决验证 gap,保留 agent forward intent
compaction validation asynchronous compaction trajectory-grounded parallel compression
核心洞察:compactor 和 agent 执行并行运行,用 judge 验证 summary 质量
📅 2025-10-01 📖 33 citations 🔗 GitHub 🔖 ID: 2510.00615
💡 peak token 减少 26-54%,小模型提升 up to 46%
agent context compression failure-driven guideline optimization dynamic context condensation
微软出品,failure-driven 压缩策略
📅 2025-09-28 📖 13 citations 🔗 GitHub 🔖 ID: 2509.23586
💡 input token 减少 39.9-59.7%,计算成本降 21.1-35.9%
trajectory reduction inference-time optimization agent waste detection LLM context pruning
通过检测 agent 轨迹中的冗余 token 来压缩上下文

计划/上下文缓存

3 篇
📅 2026-05-22 📖 citations 🔗 GitHub 🔖 ID: 2605.2063
💡 结合 Agent Workflow Memory + semantic cache
temporal semantic caching workflow pattern extraction Agent Workflow Memory
工作流模式提取 + 时序语义缓存的结合
📅 2025-09-22 📖 2 citations 🔗 GitHub 🔖 ID: 2509.1736
💡 吞吐量提升 3.6×,hit rate >85%
semantic-aware caching cross-region latency agentic LLM tool access
跨区域语义缓存,加速 tool access
📅 2025-06-17 📖 11 citations 🔗 GitHub 🔖 ID: 2506.14852
💡 cost 降低 50.31%,latency 降低 27.28%,保持 96.67% 性能
agentic plan caching task-level caching structured plan templates
测试时计划缓存,跨任务复用结构化计划

记忆压缩

7 篇
📅 2026-05-06 📖 citations 🔗 GitHub 🔖 ID: 2605.03675
💡 Acc=0.382, F1=0.412 (Qwen2.5-7B, 6GB GPU),无检索→0.382 (5%→38%)
tripartite memory episodic JSONL store five-signal retrieval PPO adaptation
三层记忆架构,OpenClaw agent runtime 原生设计,五信号加权检索引擎
📅 2026-04-14 📖 citations 🔗 GitHub 🔖 ID: 2604.07798
💡 LoCoMo 上 F1 提升 ~2.5,检索 83ms,E2E 581ms
STM/MTM/LTM三层 SLM驱动 向量粗召回 + 语义重排 固定检索预算
短/中/长期记忆三层分离,SLM 驱动轻量记忆系统
📅 2026-04-10 📖 citations 🔗 GitHub 🔖 ID: 2604.09
💡 70% 压缩下,memory retrieval 提速 1.87×,精度提升 1.0%
streaming video agent memory graph compression edge-free sampling time-decay retrieval
首个工业级流视频 agent 记忆压缩,minmax 采样 + edge-aware pruning
📅 2026-04-04 📖 citations 🔗 GitHub 🔖 ID: 2604.0167
💡 三层目录组织,用户中心上下文相关性驱动
hierarchical memory user-centric relevance primary cache user profile coupling
compact primary cache + high-priority secondary + global archive
📅 2026-04-04 📖 citations 🔗 GitHub 🔖 ID: 2604.01599
💡 LoCoMo SOTA,LongMemEval 有竞争力,无向量/图数据库
agent-native memory Context Tree AKL lifecycle zero external infra
反转记忆管道:同一个 LLM 做推理+记忆管理,层级 Context Tree
📅 2026-03-03 📖 citations 🔗 GitHub 🔖 ID: 2603.02473
💡 检索方法主导:准确率跨度 20pts (57.1%-77.2%);写策略仅 3-8pts
memory diagnostic write strategy retrieval method BM25 vs cosine vs hybrid
3×3 研究:3种写策略 × 3种检索方法。检索方法才是瓶颈
📅 2026-02-12 📖 citations 🔗 GitHub 🔖 ID: 2602.08369
💡 跨范式对齐 13 分钟完成单 GPU,优于原始 retriever <5% 训练计算
cross-paradigm memory generative subgraph retriever contrastive alignment
统一显式/参数/潜在记忆范式,两阶段训练 + 对比学习

多智能体通信剪枝

4 篇
📅 2026-03-17 📖 citations 🔗 GitHub 🔖 ID: 2603.19677
💡 信息瓶颈优化通信效率,最小化消息熵 + 动态剪枝
group-of-agents communication topology IB optimization message compression
IB (Information Bottleneck) 优化生成通信拓扑
📅 2026-03-11 📖 citations 🔗 GitHub 🔖 ID: 2603.10062
💡 将多 agent 记忆视为计算机架构问题,I/O/cache/memory 三层
shared/distributed memory three-layer hierarchy cache sharing memory consistency
位置paper,视角独特:共享vs分布式记忆,cache一致性是关键开放挑战
📅 2025-11-25 📖 citations 🔗 GitHub 🔖 ID: 2511.19969
💡 动态剪枝边连接,提升响应一致性
hierarchical graph pruning multi-modal mRAG dynamic edge pruning collaborative answering
多模态多 agent RAG,层级通信图剪枝
📅 2025-03-24 📖 citations 🔗 GitHub 🔖 ID: 2503.18891
💡 prompt token 减少 21.6%,completion token 减少 18.4%
dynamic agent role allocation communication graph pruning multi-round interaction optimization
动态消除不必要的 agent,减少多轮交互通信开销

技能/提示压缩

5 篇
📅 2026-05-26 📖 citations 🔗 GitHub 🔖 ID: 2605.28424
💡 联合优化 skill 内化 + 利用,提升分布外泛化
skill internalization OOD generalization joint optimization
skill 内部化和使用联合优化
📅 2026-05-21 📖 citations 🔗 GitHub 🔖 ID: 2605.2776
💡 像梯度下降一样优化 skill,迭代式改进
skill optimization gradient descent analogy iterative skill refinement
将 ML 的 gradient descent 概念应用于 skill 优化
📅 2026-03-31 📖 1 citations 🔗 GitHub 🔖 ID: 2603.29919
💡 description 压缩 48%,body 压缩 39%,性能提升 2.8%
skill compression progressive disclosure taxonomy-driven classification less-is-more effect
less-is-more: 删除非必要内容减少注意力分散。55,315 skills 实证研究
📅 2026-03-28 📖 citations 🔗 GitHub 🔖 ID: 2603.2785
💡 模块化 skill 封装任务特定指令和工具使用模式
agent skill code efficiency optimization modular skill units
基于 agent skill 的自动化代码效率优化
📅 2026-02-18 📖 citations 🔗 GitHub 🔖 ID: 2602.1267
💡 评估 agent skill 在不同任务上的效果
skill benchmark diverse tasks skill quality evaluation
skill 评测基准,验证压缩后 skill 质量

模型压缩对 Agent 能力影响

4 篇
📅 2026-01-18 📖 citations 🔗 GitHub 🔖 ID: 2601.12272
💡 ResNet-50 1.77G MACs 下准确率 77.04%,+0.91% over baseline
LLM-driven pruning strategy multi-agent coordination MAC constraint
LLM 驱动的剪枝策略搜索,多 agent 协调
📅 2026-01-14 📖 citations 🔗 GitHub 🔖 ID: 2601.09694
💡 LLM agent 做剪枝决策,迭代优化剪枝策略
LLM-driven pruning self-reflection agent layer-wise sensitivity adaptive compression
用 LLM agent 做自适应剪枝,结合 Wanda + gradient importance + self-reflection
📅 2025-11-19 📖 citations 🔗 GitHub 🔖 ID: 2511.15694
💡 4-bit QLoRA RL 训练产生最低量化误差网络
LRM quantization RL training 4-bit QLoRA reasoning model compression
大推理模型量化 + RL 训练的影响
📅 2025-05-26 📖 20 citations 🔗 GitHub 🔖 ID: 2505.19433
💡 4-bit 量化:规划/工具调用降 1-3%,真实应用降 10-15%
agentic capability evaluation quantization impact pruning impact 4-bit LLM agents
评估压缩对 agent 能力的影响,揭示量化/pruning 对 planning vs tool use 的不同影响

KV Cache 优化

7 篇
📅 2026-04-06 📖 citations 🔗 GitHub 🔖 ID: 2604.03143
💡 集体 KV cache 共享扩展多 agent LLM 服务
collective KV sharing multi-agent serving KV movement optimization
多 agent 间共享 KV cache,减少重复计算
📅 2026-03-17 📖 2 citations 🔗 GitHub 🔖 ID: 2603.16104
💡 端到端加速 1.56×
agentic workflow optimization LLM cache reuse query plan execution stateful operator scheduling
查询计划执行 + 有状态算子调度优化 agentic workflow
📅 2026-03-05 📖 citations 🔗 GitHub 🔖 ID: 2603.04428
💡 TTFT 降低 up to 136×(Gemma),perplexity 仅 -0.7%~+3.0%
Q4 KV cache edge device multi-agent safetensors BatchQuantizedKVCache
持久化 Q4 KV cache,Apple M4 Pro 10.2GB 缓存预算可容纳更多 agent
📅 2026-03-03 📖 citations 🔗 GitHub 🔖 ID: 2603.00188
💡 spatio-trajectory KV cache 压缩框架,component-centric simplification
training-free KV compression spatio-trajectory KV GUI agents context-aware simplification
GUI agent 的无训练 KV cache 压缩,时空轨迹压缩
📅 2026-02-26 📖 citations 🔗 GitHub 🔖 ID: 2602.22603
💡 peak token 减少 65%,仅用 215 样本训练
LRM-driven KV management auxiliary task parallel compression peak token reduction
用 Large Reasoning Model 本身做 KV 压缩,辅助任务形式并行运行
📅 2026-02-12 📖 citations 🔗 GitHub 🔖 ID: 2602.08005
💡 残差驱动 KV 压缩,利用 token 间长期相似性
residual-based KV compression inter-token similarity delta compression
发现两个规律:token 间长期相似性 + 残差压缩
📅 2025-09-23 📖 citations 🔗 GitHub 🔖 ID: 2509.17396
💡 敏感性感知的层级 KV budget 分配
episodic KV management sensitivity-aware layer-wise budget resource-constrained
长期对话的 episodic KV 缓存管理

推理加速

7 篇
📅 2026-05-27 📖 citations 🔗 GitHub 🔖 ID: 2605.21965
💡 多跳检索任务延迟降低 up to 40%
continuous speculation multi-hop retrieval lossless latency reduction theoretical framework
多跳工具使用场景的理论框架,characterize 最优可达延迟增益
📅 2026-05-18 📖 citations 🔗 GitHub 🔖 ID: 2605.1336
💡 实时交互 even with complex multi-turn tool calling
asynchronous I/O speculative tool calling real-time agent clock-based training
异步 I/O 解耦 reason-act 线程,speculative tool calling 管理不确定信息
📅 2026-04-14 📖 citations 🔗 GitHub 🔖 ID: 2604.12247
💡 自草稿 speculative decoding,层间 early exit 难度异质性
self-draft speculative decoding adaptive bounded speculation layer-wise early exit heterogeneity
自适应有界自 speculation,处理层间 early-exit 难度差异
📅 2026-04-10 📖 citations 🔗 GitHub 🔖 ID: 2604.06296
💡 同精度下最优/最差模型组合成本差距达 13-32×
client-side optimization model allocation cost-accuracy tradeoff pipeline roles
首个 client-side agent 优化框架,allocation of models to pipeline roles
📅 2026-03-19 📖 citations 🔗 GitHub 🔖 ID: 2603.18897
💡 任务完成时间降低 48.5%,工具吞吐量提升 1.8×
speculative tool execution parallel tool processing control flow pattern latency hiding
投机型工具执行,利用稳定应用层控制流 + 可预测数据依赖
📅 2026-01-23 📖 citations 🔗 GitHub 🔖 ID: 2601.22037
💡 LLM 调用减少 11.9%,任务成功率提升 4.2%
meta-tools workflow optimization redundant pattern discovery LLM call reduction
将重复工具调用序列打包为 meta-tool,减少 LLM 调用
📅 2025-10-07 📖 citations 🔗 GitHub 🔖 ID: 2510.04371
💡 游戏/电商/网页搜索环境评估,无损加速框架
speculative actions lossless acceleration faster model prediction environment steps
类比 CPU speculative execution,快模型预测未来 action 并行执行

Agent Memory 系统综述

2 篇
📅 2026-03-12 📖 citations 🔗 GitHub 🔖 ID: 2603.11768
memory governance temporal decay structuring impact retrieval effectiveness
记忆治理:风险+机制+评估,structuring 对检索有效性的影响
📅 2026-03-08 📖 12 citations 🔗 GitHub 🔖 ID: 2603.0767
💡 五类机制:context-resident compression, RAG, reflective, hierarchical, policy-learned
agentic memory survey policy-learned management hierarchical virtual context RAG
全面梳理 agent 记忆中持久化的信息,1136 references 分析

Experience Compression 综述

4 篇
📅 2026-04-17 📖 citations 🔗 GitHub 🔖 ID: 2604.15877
💡 5-20× episodic memory, 50-500× procedural skills, 1000×+ declarative rules
experience compression spectrum memory-skills-rules unification missing diagonal cross-level adaptive compression
⭐ 核心洞察:20+ 系统都在固定压缩级别运作,无自适应跨级压缩(missing diagonal)。memory 和 skills 社区互相不引用(<1% cross-citation)
📅 2026-03-31 📖 citations 🔗 GitHub 🔖 ID: 2603.29678
💡 通过 view 导向编译实现 agent 历史的自我压缩
conversation compiler optical self-compression agent trace analysis view-oriented
Agentocr 引用:光学校自压缩
📅 2026-03-31 📖 citations 🔗 GitHub 🔖 ID: 2603.29193
💡 解决上下文衰减和响应一致性下降问题
adaptive context compression long-running interaction context decay persistent conversation
长程交互中的自适应上下文压缩技术
📅 2025-07-22 📖 citations 🔗 GitHub 🔖 ID: 2507.16784
💡 TIM-8b 递归分解任务,自主管理上下文,性能匹配复杂 agent 实现
subconscious threads recursive subtask decomposition working memory pruning TIM-8b
用递归子任务分解 + 工作记忆剪枝替代显式记忆系统

Coding Agent 特化压缩

3 篇
📅 2026-05-15 📖 citations 🔗 GitHub 🔖 ID: 2605.15315
💡 多轮任务节省 31% tokens,单轮 Exact Match +3.5
multi-rubric CRF semantic evidence dependency support mixture-of-experts gating
分解代码相关性为 semantic evidence + dependency support 两个维度,双 CRF + MoE gating
📅 2026-02-18 📖 citations 🔗 GitHub 🔖 ID: 2602.12852
💡 工具调用轮次减少约 20%,同时提升准确率
web agent pruning DAG mining state graph tool-call reduction 20%
图结构建模搜索过程,最小必要 DAG 挖掘,保留 essential reasoning
📅 2026-01-23 📖 citations 🔗 GitHub 🔖 ID: 2601.16746
💡 agent 任务 token 减少 23-54%(甚至成功率提升),单轮 14.84× 压缩
coding agent pruning 0.6B neural skimmer self-adaptive token reduction 23-54%
轻量神经网络 skimmer 动态选择相关代码行,针对 coding agent 特化

System Co-Design

4 篇
📅 2026-04-22 📖 citations 🔗 GitHub 🔖 ID: 2604.20938
💡 自动化 harness 配置搜索,取代手动 stacking
harness optimization Bayesian optimization flag-gated harness automated configuration
将 harness design 作为 first-class ML problem,constrained noisy BO over mixed-variable config space
📅 2026-04-09 📖 citations 🔗 GitHub 🔖 ID: 2604.07857
💡 能效 agentic AI 综述,统一 taxonomy:模型简化 + 计算控制 + 跨层协同设计
energy-efficient agentic AI network-aware cross-layer co-design wireless edge
网络感知能效优化,面向无线边缘资源受限部署
📅 2026-03-11 📖 citations 🔗 GitHub 🔖 ID: 2603.10342
💡 单 GPU 并发 tool-augmented SLM agent 服务,算法-系统协同设计
consumer GPU serving scheduling co-optimization resource management concurrent tool-augmented SLM
消费级 GPU 上的 agent serving,调度和资源管理协同优化
📅 2026-02-20 📖 citations 🔗 GitHub 🔖 ID: 2512.18337
💡 四组件协同实现:协作减少冗余计算,压缩提升缓存利用,调度优化 GPU,speculative 加速
algorithm-system co-design cognitive guidance predictive token speculation adaptive compression
系统级协同:认知引导 + 预测性 token 推测 + 动态调度 + 自适应压缩

检索/RAG 压缩

2 篇
📅 2026-02-22 📖 citations 🔗 GitHub 🔖 ID: 2602.20426
💡 100+ candidate tools 上未见工具的跨域泛化能力
tool description rewriting Trace-Free+ curriculum learning cross-domain generalization
优化 tool interface 而非 agent,冷启动友好,课程学习从 trace-rich 到 trace-free
📅 2026-01-26 📖 citations 🔗 GitHub 🔖 ID: 2601.18886
💡 多语言零成本上下文剪枝
multilingual context pruning zero-cost RAG cross-lingual
多语言 RAG 场景的上下文剪枝