64 KiB
Wiki Log
所有 wiki 操作的按时间顺序记录。仅追加。 格式:
## [YYYY-MM-DD] action | subject操作类型:ingest, update, query, lint, create, archive, delete 当此文件超过 500 条记录时,轮换:重命名为 log-YYYY.md,重新开始。
2026-06-25 — ingest | Wan-Streamer v0.1 (arXiv:2606.25041, 2026)
- 添加论文 wan-streamer: "Wan-Streamer v0.1: End-to-end Real-time Interactive Foundation Models" — 阿里巴巴 Wan Team 的端到端流式全双工音视频交互基础模型
- 新增 5 个概念页: block-causal-attention, full-duplex-interaction, thinker-performer-pipeline, causal-multimodal-vae, end-to-end-streaming-interaction
- 更新 4 个已有概念页: flow-matching, kv-cache, diffusion-transformer, native-streaming-ar-training
- 来源: https://arxiv.org/abs/2606.25041
2026-06-25 — ingest | Large Language Gibbs (arXiv:2606.19264, 2026)
- 添加论文 large-language-gibbs: "Structured Inference with Large Language Gibbs" — Edinburgh 团队的 LLM + Gibbs 采样结构化概率推断框架
- 新增 5 个概念页: llm-mcmc, barker-gibbs, gambling-gibbs, order-bias-removal, llm-consistent-reasoning
- 来源: https://arxiv.org/abs/2606.19264
2026-06-25 — ingest | Latent CoT Supervision (arXiv:2606.20075, ICML 2026)
- 添加论文 latent-cot-supervision: "What Makes Effective Supervision in Latent Chain-of-Thought: An Information-Theoretic Analysis" — 从信息论角度分析潜推理的有效监督机制
- 新增 7 个概念页: dual-collapse, trajectory-supervision, space-supervision, unified-latent-probe, information-performance-binding, generative-reconstruction-latent, geometric-compression-latent
- 来源: https://arxiv.org/abs/2606.20075
[2026-06-25] create | Agent Memory Five-Category Model (sz 记忆架构设计)
- 新增概念 agent-memory-five-category-model: sz 五类记忆模型——知识/概念/Cron/用户绑定/前瞻记忆的完整分类与 Atlas 映射
- 新增概念 prospective-memory-index: 前瞻记忆索引——第 5 类记忆(计划/想法/洞察)的锚点设计:语义关联衰减、LLM 重要性分类器、闭合状态管理
- 更新 atlas-memory-system: 添加五类模型扩展与前瞻索引交叉引用
[2026-06-25] ingest | LongMemEval: Benchmarking Long-Term Interactive Memory (arXiv:2410.10813, UCLA/Tencent, ICLR 2025)
- 添加论文 longmem-eval-2025: "LongMemEval: Benchmarking Chat Assistants on Long-Term Interactive Memory" — 500 题 × 5 能力记忆基准 + 三阶段统一框架
- 新增 5 个概念页: long-term-interactive-memory, longmem-eval, memory-indexing-retrieval-reading, fact-augmented-key-expansion, time-aware-query-expansion
- 来源: https://arxiv.org/abs/2410.10813
[2026-06-25] ingest | Engram: Conditional Memory via Scalable Lookup (arXiv:2601.07372, PKU/DeepSeek-AI)
- 添加论文 engram-conditional-memory-2026: "Conditional Memory via Scalable Lookup" — 条件记忆作为 MoE 的互补稀疏轴
- 新增 5 个概念页: conditional-memory, engram, sparsity-allocation, ngram-embedding, memory-compute-decoupling
- 来源: https://arxiv.org/abs/2601.07372
[2026-06-24] ingest | Arbor: Hypothesis-Tree Refinement (arXiv:2606.11926, RUC/MSR)
- 添加论文 arbor-htr-2026: "Toward Generalist Autonomous Research via Hypothesis-Tree Refinement" — Coordinator+Executor 架构 + 假设树持久化 + AO 形式化
- 新增 5 个概念页: hypothesis-tree-refinement, coordinator-executor-architecture, autonomous-optimization-ao, insight-backpropagation, research-hypothesis-tree
- 来源: https://arxiv.org/abs/2606.11926
[2026-06-24] ingest | ME² + TRM: Complex Reasoning Optimization (arXiv:2602.08498, ICML 2026)
- 添加论文 me2-trm-reasoning-2026: "Characterizing, Evaluating, and Optimizing Complex Reasoning" — ME² 原则 + DAG 推理建模 + Thinking Reward Model
- 新增 5 个概念页: me2-principle, thinking-reward-model, dag-reasoning-evaluation, trm-preference-dataset, reasoning-quality-optimization
- 复用: large-reasoning-models, reward-model, grpo
- 来源: https://arxiv.org/abs/2602.08498
[2026-06-24] ingest | Atlas Agent 记忆系统架构(公众号技术文章)
- 添加文章 atlas-agent-memory-architecture-2026: "Atlas Agent 记忆系统架构" — 三索引分型 + BM25+dense 混合召回 + consolidation + soft-supersession
- 新增 8 个概念页: atlas-memory-system, agent-memory-taxonomy, hybrid-recall-pipeline, verbatim-pre-recall, memory-consolidation, soft-supersession, per-index-time-decay, gbrain-memory
- 复用: bm25-financial-retrieval
- 来源: https://mp.weixin.qq.com/s/fypjVWJBQg_MZV9OMfPpIA
[2026-06-24] ingest | VLA-JEPA (arXiv:2602.10098, cs.RO/cs.CV)
- 添加论文 vla-jepa-2026: "VLA-JEPA" — JEPA 范式引入 VLA,leakage-free state prediction 修复 latent-action 预训练的信息泄漏
- 新增 7 个概念页: vla-jepa, leakage-free-state-prediction, latent-world-model, latent-action-pretraining, information-leakage-vla, jepa-for-robotics, appearance-bias-vla
- 复用: jepa, vla-vision-language-action, world-model-lecun, flow-matching
- 来源: https://arxiv.org/abs/2602.10098
[2026-06-24] ingest | The Personalization Trap (arXiv:2510.09905, cs.AI/cs.CL, Amazon)
- 添加论文 personalization-trap-2025: "The Personalization Trap" — 用户记忆如何系统性改变 LLM 情感推理,优势画像比劣势画像获更准确情感解读
- 新增 7 个概念页: personalization-trap, user-memory-bias, emotional-reasoning-bias, social-capital-framework, situational-test-emotional-understanding, intersectional-persona-evaluation, persona-invariant-reasoning, dpo-bias-mitigation
- 复用: dpo
- 来源: https://arxiv.org/abs/2510.09905
[2026-06-24] ingest | Unlimited OCR Works (arXiv:2606.23050, cs.CV/cs.CL, Baidu)
- 添加论文 unlimited-ocr-works-2026: "Unlimited OCR Works" — R-SWA 注意力机制实现恒定 KV cache 的一次前向长程 OCR
- 新增 10 个概念页: reference-sliding-window-attention, constant-kv-cache, long-horizon-parsing, deepseek-ocr, deepencoder, omnidocbench, end-to-end-ocr, unlimited-ocr, megatron-lm, sglang
- 来源: https://arxiv.org/abs/2606.23050
[2026-06-24] ingest | 金融行业大模型落地实践(林金曙,DAcon 2026)
- 添加文章 financial-llm-practice-2026: "金融行业大模型落地实践:从长文档检索到 Agent 工程" — 恒生电子金融 LLM 工程实践全链路分享
- 新增 9 个概念页: pageindex, agentic-rag, financial-llm-requirements, financial-llm-model-selection, bm25-financial-retrieval, agent-skill-atomization, financial-agent-permission, aidb, financial-llm-deployment
- 来源: https://mp.weixin.qq.com/s/3iObkj6BKhZzphJ1URVOKg
[2026-06-23] ingest | DCGWM: Dual-Channel Grounded World Modeling (arXiv:2606.18688, cs.LG/cs.AI 2026)
- 添加论文 hazare-dcgwm-2026: "结构防止目标干扰坍缩的双通道接地世界建模" — 识别 OIC 新失效模式,提出分区潜在空间+内向梯度流架构
- 新增 6 个概念页: objective-interference-collapse, dcgwm, inward-only-gradient-flow, asymmetric-grounding-adherence-loss, rollout-drift, isolation-necessity-theorem
- 复用已有概念: jepa, vicreg, world-models-rl, representation-collapse
- 来源: https://arxiv.org/abs/2606.18688
- 注: Position paper, 实验验证进行中
[2026-06-23] ingest | Fisher Width: A Geometric Measure of Complexity on Statistical Manifolds (arXiv:2606.18306, cs.LG/stat.ML 2026)
- 添加论文 vu-fisher-width-2026: "统计流形上的几何复杂度度量" — 将 Gaussian width 推广到 Fisher 几何,引入 Fisher width 及其泛化界
- 新增 6 个概念页: fisher-width, gaussian-width, statistical-manifold, fisher-lipschitz, lifting-identity, empirical-fisher
- 复用已有概念: fisher-information-metric, information-geometry, generalization-bounds, natural-gradient-descent
- 来源: https://arxiv.org/abs/2606.18306
[2026-06-23] ingest | A Bifurcation Theory Framework for GD on the Edge of Stability (arXiv:2606.15551, cs.LG 2026)
- 添加论文 gan-bifurcation-eos: "分岔理论框架下的梯度下降稳定边缘分析" — 将 EoS 稳定性归结为 flip 分岔的 c₁ 符号,统一乘积稳定性为特例
- 新增 8 个概念页: edge-of-stability, flip-bifurcation, first-lyapunov-coefficient, manifold-of-minimizers, normal-tangent-decomposition, sharpness, product-stability, center-manifold-theorem
- 来源: https://arxiv.org/abs/2606.15551
[2026-06-22] ingest | NANO Filter — Nonlinear Bayesian Filtering with Natural Gradient Gaussian Approximation (arXiv:2410.15832, eess.SY)
- 添加论文 nano-filter: "NANO 自然梯度高斯近似滤波" — 跳出线性化框架,直接在 Gaussian 流形上优化后验
- 新增 11 个概念页: bayesian-filtering, kalman-filter, natural-gradient-descent, gaussian-filtering, stein-lemma, gibbs-posterior, gaussian-manifold, moment-matching-filter, pseudo-huber-loss, posterior-linearization-filter, nano-filter
- 1 个 Review: nano-filter-20260622
- 来源: https://arxiv.org/abs/2410.15832
[2026-06-21] ingest | Jordan — A Collectivist, Economic Perspective on AI (arXiv:2507.06268, cs.CY)
- 添加论文 jordan-collectivist-ai-2025: "AI 的集体主义经济学视角" — LLM 作为集体主义制品,三种思维方式的融合
- 新增 4 个概念页: statistical-contract-theory, e-values, data-markets, probability-matching
- 1 个 Review: jordan-collectivist-ai-review-20260621
- 更新已有概念: collectivist-ai (追加案例), prediction-driven-inference (追加 PPI 学术溯源)
- 来源: https://arxiv.org/abs/2507.06268
[2026-06-21] ingest | Michael I. Jordan MLST访谈 — 机器之心编译
- 添加文章 michael-jordan-mlst-collectivist-ai-2026: "AI 的集体主义经济学与虚假的 AGI 二元论"
- 新增 6 个概念页: collectivist-ai, uncertainty-taxonomy, prediction-driven-inference, foundation-model-frontier-bias, anthropomorphization-critique, agi-critique
- 更新已有概念: uncertainty-quantification — 追加 Jordan 社会-经济扩展维度
- 来源: https://mp.weixin.qq.com/s/VEo23R0yst6wjdyzVicYUQ (arXiv:2507.06268)
[2026-06-21] ingest | LLM沉默螺旋综述 — 数据派THU (李媛媛)
- 添加文章 llm-spiral-of-silence-2026: "大模型沉默螺旋:当算法催生数字从众" — 系统性综述 LLM 在 RAG 闭环与多智能体交互中的算法驱动沉默螺旋
- 新增 12 个概念页: spiral-of-silence, pretraining-statistical-bias, context-anchoring, role-setting-entrenchment, rlhf-alignment-amplification, rag-closed-loop, multi-agent-spiral, content-homogenization, information-cocoons, content-diversity-decay, opinion-polarization, temperature-sampling
- 更新已有概念: rlhf, rag — 追加沉默螺旋维度交叉引用
- 来源: https://mp.weixin.qq.com/s/ZKrx4BzmiOUBsfPVY9YHyw
[2026-06-20] ingest | MaineCoon (arXiv:2606.17800, cs.CV)
- 添加论文 mainecoon: "MaineCoon: Pursuing A Real-Time Audio-Visual Social World Model" — 首个实时流式音视频社交世界模型 (22B, 47.5 FPS)
- 新增 15 个概念页: social-world-model, self-resampling, reinforced-online-policy-distillation, agentic-streaming-inference, agentic-cache-manager, look-ahead-buffer-controller, forward-repair-ladder, socialvideo-bench, audio-visual-representation-alignment, domain-aware-preference-optimization, audio-visual-generation, autoregressive-video-generation, streaming-generation, diffusion-transformer, social-video, drifting
- 1 个 Review: mainecoon-review-20260620
- 来源: https://arxiv.org/abs/2606.17800
[2026-06-19] ingest | ACE-Router (arXiv:2601.08276, cs.AI)
- 添加论文 yao-ace-router-2026: "ACE-Router: Generalizing History-Aware Routing from MCP Tools to the Agent Web" — 训练专用路由器
- 新增 7 个概念页: ace-router, history-aware-routing, candidate-graph, self-evolutionary-mutation, trajectory-synthesis, light-routing-agent, agent-web
- 来源: https://arxiv.org/abs/2601.08276
[2026-06-19] ingest | Dynamic ReAct (arXiv:2509.20386, cs.SE)
- 添加论文 gaurav-dynamic-react-2025: "Dynamic ReAct: Scalable Tool Selection for Large-Scale MCP Environments" — 五架构→Search and Load 最优
- 新增 6 个概念页: dynamic-react, meta-tools, search-and-load, context-enriched-embeddings, default-tools, tool-registry
- 来源: https://arxiv.org/abs/2509.20386
[2026-06-19] ingest | MCP-Zero (arXiv:2506.01056, cs.AI)
- 添加论文 fei-mcp-zero-2025: "MCP-Zero: Active Tool Discovery for Autonomous LLM Agents" — 主动工具发现范式
- 新增 6 个概念页: active-tool-discovery, active-tool-request, hierarchical-semantic-routing, iterative-capability-extension, mcp-protocol, mcp-tools-dataset
- 来源: https://arxiv.org/abs/2506.01056
[2026-06-19] ingest | MemOS Agent 记忆基础设施(熊飞宇/MemTensor, DataFun)
- 添加文章 memtensor-memos-agent-memory-2026: MemOS 记忆系统从效率工具到生存关键
- 新增 9 个概念页: agent-memory-system, layered-memory-architecture, model-driven-vs-app-driven-memory, mem2skill, memory-governance, clawforce, agent-memory-lifecycle, memcube, memory-dedup-pipeline
- 来源: https://mp.weixin.qq.com/s/5Wo91nzstNtCIV9chnuQmw
[2026-06-19] ingest | Six Choices Every AI Engineer Has to Make (Nobrega, 数据派THU)
- 添加文章 nobrega-ai-production-tradeoffs-2026: AI 工程师的 6 种生产权衡
- 新增 9 个概念页: ai-production-tradeoffs, build-vs-buy-llm, cace-principle, ml-technical-debt, data-quality-vs-quantity, batch-vs-real-time-inference, prompt-engineering-vs-fine-tuning, human-in-the-loop, selective-hitl, data-swamp
- 来源: https://mp.weixin.qq.com/s/GESoyR0qpxP4fPtHZjonKA
[2026-06-19] ingest | Agent Skills Survey (arXiv:2605.07358)
- 添加论文 zhou-agent-skills-survey-2026: "A Comprehensive Survey on Agent Skills: Taxonomy, Techniques, and Applications" — agent skill 生命周期的系统性综述
- 新增 12 个概念页: agent-skill, procedural-gap, skill-lifecycle, skill-representation, skill-acquisition, skill-retrieval, skill-selection, skill-evolution, skill-composition, agent-skill-ecosystem, passive-vs-active-knowledge, runtime-governance
- 来源: https://arxiv.org/abs/2605.07358
[2026-06-18] ingest | Transformers are SSMs: Structured State Space Duality (arXiv:2405.21060, ICML 2024)
- 添加论文 dao-transformers-are-ssms-2024: "Transformers are SSMs" — Dao & Gu 提出 SSD 框架统一 SSM 和 Attention,设计 Mamba-2 架构 (2-8x 加速)
- 新增 9 个概念页: structured-state-space-duality, semiseparable-matrices, structured-masked-attention, mamba-2, ssd-algorithm, linear-attention, selective-state-space-models, tensor-contraction-duality, head-structure-ssm
- 更新已有: mamba-ssm, state-space-models — 添加 Mamba-2 反向链接
- 来源: https://arxiv.org/abs/2405.21060
[2026-06-18] ingest | Thinking-Based Non-Thinking (arXiv:2601.04805, Preprint)
- 添加论文 gan-thinking-based-non-thinking-2026: "Thinking-Based Non-Thinking" — TNT: 利用思考模式 solution 长度动态限制非思考 token,解决混合推理模型的 Reward Hacking
- 新增 10 个概念页: hybrid-reasoning-models, reward-hacking, overthinking, thinking-based-non-thinking, dynamic-token-limit, non-thinking-mode, thinking-mode, ellipsis-prompt, large-reasoning-models, token-level-policy-gradient
- 来源: https://arxiv.org/abs/2601.04805
[2026-06-18] ingest | RWKV-7 "Goose" with Expressive Dynamic State Evolution (arXiv:2503.14456)
- 添加论文 peng-rwkv7: "RWKV-7 Goose" — 广义 Delta 规则 + 向量值门控,首个超越 TC^0 的并行化 RNN
- 新增 8 个概念页: rwkv, delta-rule, generalized-delta-rule, vector-valued-gating, in-context-learning-rate, dynamic-state-evolution, token-shift, wkv-time-mixing, regular-language-recognition
- 更新: enhanced-state-space-models (扩充 RWKV-7 小节)
- 新增 review: rwkv7-review-20260618
- 来源: https://arxiv.org/abs/2503.14456 | 代码: https://github.com/RWKV/RWKV-LM
[2026-06-18] ingest | Mamba: Linear-Time Sequence Modeling with Selective State Spaces (arXiv:2312.00752)
- 添加论文 gu-mamba: "Mamba" — 选择性状态空间模型,线性时间序列建模
- 新增 7 个概念页: selective-state-space, hardware-aware-algorithm, structured-state-space-models, content-based-reasoning, selective-copy, induction-heads, hippo
- 更新 2 个已有概念页: mamba-ssm (大幅扩充), state-space-models (追加论文引用)
- 新增 review: mamba-review-20260618
- 来源: https://arxiv.org/abs/2312.00752 | 代码: https://github.com/state-spaces/mamba
[2026-06-18] ingest | Predicting Future Utility: Global Combinatorial Optimization for Task-Agnostic KV Cache Eviction (arXiv:2602.08585, ICML 2026)
- 添加论文 tang-lukv: "LU-KV" — 基于全局组合优化的 head 级 KV Cache 预算分配框架
- 新增 18 个概念页: kv-cache, kv-cache-eviction, lukv, oracle-importance, optimality-gap, long-horizon-utility, marginal-utility, global-combinatorial-optimization, convex-hull-relaxation, offline-profiling, head-level-budget-allocation, intra-head-eviction, cross-head-budget-allocation, heuristic-metric, snapkv, pyramidkv, adkv, keydiff
- 新增 review: lukv-review-20260618
- 来源: https://arxiv.org/abs/2602.08585
[2026-06-18] ingest | The Topological Trouble With Transformers (arXiv:2604.17121, Preprint)
- 添加论文 mozer-topological-trouble-transformers-2026: "The Topological Trouble With Transformers" — 分析前馈 Transformer 状态追踪的拓扑性局限并提出循环架构分类法
- 新增 16 个概念页: state-tracking, feedforward-depth-limitation, belief-state, depth-dilemma, recurrent-transformer-architectures, recurrence-taxonomy, depth-recurrence, step-recurrence, coarse-grained-recurrence, latent-thought-models, attractor-dynamics, enhanced-state-space-models, representational-alignment, sequential-dependency, autoregressive-unrolling, state-space-models
- 来源: https://arxiv.org/abs/2604.17121
[2026-06-17] ingest | Uncertainty Estimation and Generalization Bounds for Modern Deep Learning (PhD Thesis, arXiv:2606.13818, cs.LG 2026)
- 添加论文 ortega-phd-thesis: "Uncertainty Estimation and Generalization Bounds" — PhD论文,DVIP + VaLLA + FMGP + PAC-Chernoff泛化界
- 新增 10 个概念页: deep-variational-implicit-process, variational-linearized-laplace-approximation, fixed-mean-gaussian-process, pac-bayesian-bounds, implicit-processes, function-space-modeling, generalization-bounds, double-descent, deep-gaussian-process, gaussian-process
- UAM 博士论文,统一 Bayesian 方法 + PAC-Bayesian 理论 + 大偏差分析
- 来源: https://arxiv.org/abs/2606.13818
[2026-06-17] ingest | Learning to Adapt: Representation-Based RL for Multi-Task Skill Transfer (arXiv:2606.12890, cs.RO 2026)
- 添加论文 repmt-sac: "RepMT-SAC" — 谱 MDP 分解 + 上游-下游两阶段学习的多任务 SAC,四旋翼跟踪 +30%
- 新增 8 个概念页: rep-mt-sac, spectral-mdp-decomposition, task-invariant-representation, task-conditioned-policy, quadrotor-trajectory-following, upstream-downstream-learning, soft-actor-critic, task-distribution
- Harvard SEAS + MIT,IsaacSim 验证,零样本 ID + 少样本 OOD
- 来源: https://arxiv.org/abs/2606.12890
[2026-06-17] ingest | Weighted Universal Approximation of Differentiable Maps on Infinite-Dimensional Manifolds (arXiv:2606.09820, math.FA 2026)
- 添加论文 weighted-uat-manifolds: "Weighted UAT" — 无限维流形上 FNN 的加权通用逼近,含导数
- 新增 8 个概念页: functional-input-neural-networks, universal-approximation-theorem, nachbin-theorem, weighted-spaces, infinite-dimensional-manifolds, bastiani-calculus, non-anticipative-functionals, signature
- 77页 math.FA 核心论文,首次将 UAT 从紧集扩展到加权非紧空间并包含导数逼近
- 来源: https://arxiv.org/abs/2606.09820
[2026-06-17] ingest | Bellman–Taylor Score Decoding for MDPs with State-Dependent Feasible Action Sets (arXiv:2606.10979, cs.AI 2026)
- 添加论文 bellman-taylor-score-decoding: "Bellman–Taylor Score Decoding" — Taylor 展开 Q 函数将约束 MDP 映射为潜在得分 MDP,标准 DRL 直接可用
- 新增 8 个概念页: bellman-taylor-score-decoding, latent-score-mdp, state-dependent-feasible-action-sets, action-decoder, post-action-configuration, taylor-expansion-q-function, queueing-network-control, btsd-ppo, continuation-value-function
- HKUST IEDA,排队网络控制验证,不需求导解码器,性能保证可分解为近似误差+学习误差
- 来源: https://arxiv.org/abs/2606.10979
[2026-06-17] ingest | A Geometric View for Understanding Concept Learning and Neuron Interpretation in Sparse Autoencoders (arXiv:2606.07007, cs.LG 2026)
- 添加论文 geometric-sae-concepts: "A Geometric View" — SAE 概念学习与神经元解释的统一几何框架,集合论 + 形式概念分析
- 新增 12 个概念页: sparse-autoencoder, polysemanticity, mechanistic-interpretability, formal-concept-analysis, concept-learning, feature-splitting, feature-absorption, feature-family, absolute-gating, hyperplane-arrangements, concept-lattice, superposition
- UW Paul G. Allen School,区分 concept detection / separation / approximation 三层学习,建立概念格组织多对多关系
- 来源: https://arxiv.org/abs/2606.07007
[2026-06-17] ingest | From Ticks to Flows: Dynamics of Neural RL in Continuous Environments (ICLR 2026, arXiv:2606.04275, cs.LG)
- 添加论文 ticks-to-flows: "From Ticks to Flows" — 连续时间 RL 的双时间尺度理论分析,SDE + NTK + 鞅 CLT
- 新增 12 个概念页: continuous-time-rl, stochastic-differential-equation, wiener-process, ito-calculus, two-time-scale-process, exploratory-dynamics, linearized-neural-network, infinite-width-limit, neural-tangent-kernel, martingale-clt, linear-quadratic-regulator, control-affine-mdp
- ICLR 2026 接收,Brown University,首次给出连续 RL 中 NN 参数梯度更新的状态分布演化方程
- 来源: https://arxiv.org/abs/2606.04275
[2026-06-17] ingest | TARPO: Token-Wise Latent-Explicit Reasoning via Action-Routing Policy Optimization (arXiv:2606.05859, cs.CL 2026)
- 添加论文 tarpo: "TARPO" — 纯 RL 驱动的逐 token 潜在-显式混合推理框架,自适应 hard/soft 切换
- 新增 12 个概念页: latent-reasoning, coconut, soft-token, hard-token, hybrid-reasoning, hrpo, token-wise-routing, action-routing-policy, action-head-router, reparameterization-exploration, gumbel-softmax, continuous-representation
- 来自南开大学 TMCC,Qwen2.5 (1.5B-7B) 和 Llama-3.1-8B 验证
- 来源: https://arxiv.org/abs/2606.05859
[2026-06-16] ingest | Advances in Temporal Point Processes: Bayesian, Neural, and LLM Approaches (TMLR, 2026 OpenReview: SXgGKkShhT)
- 添加论文 advances-temporal-point-processes-2026: "Advances in Temporal Point Processes" — TPP 综述,首篇同时覆盖 Bayesian/Neural/LLM 三大范式
... [OUTPUT TRUNCATED - 538 chars omitted out of 50538 total] ...
, lifecycle-orchestration, observability, verification-evaluation, governance-security, cost-quality-speed-trilemma, capability-control-tradeoff, harness-coupling-problem, binding-constraint-thesis, prompt-to-harness-evolution, trace-native-evaluation, standard-agent-handoffs, adaptive-harness-simplification, hardening-execution-environments, reliable-state-long-running-agents, context-state-estimation, agent-frameworks-to-platforms
- 来源: 用户上传 PDF(用户 o9cq80wQvcn_qxHaHlEso2Bn3qoU@im.wechat)
- Wiki 规模: 373 → 395 页
[2026-05-21] ingest | KORE (arXiv:2510.19316, ICML 2026)
- 添加论文 kore-knowledge-injection: "KORE: Enhancing Knowledge Injection via Knowledge-Oriented Controls" — 知识导向控制协同方法,零空间投影+知识树实现适应与保留双优
- 新增 6 个概念页: kore-augmentation, kore-constraint, knowledge-tree, null-space-projection-knowledge, covariance-matrix-knowledge, hars
- KORE 是 MMEVOKE 系列工作的解决方案论文,同一作者团队
- 来源: https://arxiv.org/abs/2510.19316
[2026-05-21] ingest | When Large Multimodal Models Confront Evolving Knowledge (arXiv:2505.24449, ICLR 2026)
- 添加论文 when-large-multimodal-models-confront-evolving-knowledge: "When Large Multimodal Models Confront Evolving Knowledge" — 多模态进化知识注入首个基准MMEVOKE,揭示双重挑战并探索知识增强与保留方案
- 新增 12 个概念页: evolving-knowledge-injection, mme-voke, knowledge-aware-augmentation, knowledge-agnostic-augmentation, capability-degradation, data-replay, moe-lora, multimodal-rag, knowledge-retention, knowledge-adaptation, self-evolving-benchmark, sufficient-context-paradox
- 来源: https://arxiv.org/abs/2505.24449
2026-05-15 — ingest | Continuous Thought Machines (arXiv:2505.05522, NeurIPS 2025)
- 添加论文 darlow-ctm-2025: "Continuous Thought Machines" — 以神经同步为表示的新型架构,NLMs + Neural Synchronization 两大创新
- 新增 11 个概念页: continuous-thought-machine, neuron-level-models, neural-synchronization, internal-ticks, synapse-model, certainty-based-loss, adaptive-computation-time, internal-world-model, neuron-pairing, temporal-decay-neural, pre-activation-history
- 核心创新: 每个神经元私有 NLM 替代统一激活函数 + 激活历史内积作为同步表示
- 实验亮点: 迷宫泛化(39×39→99×99)、ImageNet 原生自适应计算、Parity 可解释策略
- 作者含 Llion Jones (Attention Is All You Need 合著者), 机构: Sakana AI
- 来源: https://arxiv.org/abs/2505.05522
2026-05-15 — ingest | NeurIDA (arXiv:2512.08483v3, cs.DB 2025)
- 添加论文 zeng-neurida-2025: "NeurIDA: Dynamic Modeling for Effective In-Database Analytics" — 端到端自主库内分析系统,通过动态装配定制模型解决 ML 静态性与 RDBMS 动态性的范式鸿沟
- 新增 15 个概念页: neurida, dynamic-in-database-modeling, dime-dynamic-in-database-modeling-engine, composable-base-model-architecture, query-intent-analyzer, conditional-model-dispatcher, zero-cost-proxies, dynamic-relation-modeling, dynamic-model-fusion, data-slice, base-table-embedding, in-database-analytics, relational-graph, analytical-report-synthesizer, tabular-foundation-models
- 核心创新: DIME 四阶段管线(表嵌入→关系建模→上下文融合→预测),从共享组件在查询时动态装配定制模型
- 实验: 5 数据集 10 任务,AUC-ROC ↑12%, MAE ↓25%, 延迟开销仅 1.1×–2.1×
- 来源: https://arxiv.org/abs/2512.08483v3
2026-05-12 — ingest | TBA (arXiv:2503.18929, NeurIPS 2025)
- 添加论文 bartoldson-tba-2025: "Trajectory Balance with Asynchrony" — GFlowNet TB 目标 × 异步分布式 RL
- 新增 8 个概念页: tba, trajectory-balance-objective, asynchronous-rl-llm, off-policy-llm-post-training, gflownet-fine-tuning, replay-buffer-rl-llm, searcher-trainer-decoupling, reward-recency-sampling
- 核心创新: 利用 TB 目标的 off-policy 兼容性,实现 Searcher-Trainer 解耦,4×–50× 训练加速
- TBA′ 在高度 off-policy 设置下超越 Dr. GRPO(MATH, Qwen 2.5 7B)
- 来源: https://arxiv.org/abs/2503.18929 | 代码: https://github.com/bbartoldson/TBA
2026-05-12 — ingest | MathForge (arXiv:2601.20614, ICLR 2026)
- 添加论文 dai-mathforge-2026: "Harder Is Better" — 难度感知 GRPO + 多维度问题改写
- 新增 8 个概念页: grpo, mathforge, dgpo, dgae, dqw, mqr, update-magnitude-imbalance, math-question-reformulation
- 核心发现: GRPO 存在更新幅度难度不平衡 (Theorem 1), DGAE 用 MAD 替代 std 解决 (Theorem 2)
- MQR 三维改写策略: Background (99%), Term (97%), Sub-Problem (97%) 答案保持率
- 来源: https://arxiv.org/abs/2601.20614 | 代码: https://github.com/AMAP-ML/MathForge
[2026-05-18] ingest | Pre-train Space Reinforcement Learning (arXiv:2604.14142, 2026)
- 添加论文 pre-train-space-reinforcement-learning: "Pre-train Space RL" — 在预训练空间中应用 RL,NSR-PreRL 剪枝错误路径并激发内生推理,DSRL 全面超越 GRPO
- 新增 11 个概念页: pre-train-space-reinforcement-learning, dual-space-rl, post-train-space-rl, negative-sample-reinforcement, positive-sample-reinforcement, gradient-alignment, policy-reincarnation, endogenous-reasoning, shared-parameter-influence, distribution-shift, on-policy-learning-collapse
- 来源: https://arxiv.org/abs/2604.14142
[2026-05-14] ingest | StreamingLLM: 基于注意力汇的高效流式语言模型 (arXiv:2309.17453, ICLR 2024)
- 添加论文 streaming-llm: "Efficient Streaming Language Models with Attention Sinks" — 发现 Attention Sink 现象,提出无需微调的无限长流式推理框架
- 新增 5 个概念页: length-extrapolation, rolling-kv-cache, sink-token, softmax-off-by-one, window-attention
- 更新概念 attention-sinks: 从占位符扩展为完整内容(含数学推导、实验证据、应用)
- 来源: https://arxiv.org/abs/2309.17453
- 创建 5 个作者实体页: guangxuan-xiao, yuandong-tian, beidi-chen, song-han, mike-lewis
[2026-05-14] ingest | LLMs Corrupt Your Documents When You Delegate (arXiv:2604.15597, April 2026)
- 添加论文 laban-llms-corrupt-documents-delegate: "LLMs Corrupt Your Documents When You Delegate" — DELEGATE-52 基准揭示LLM在委托工作中静默破坏文档
- 新增 11 个概念页: delegate-52, backtranslation-round-trip-relay, round-trip-reconstruction-score, document-degradation, critical-failures, delegated-work, long-horizon-evaluation, semantic-equivalence, domain-specific-evaluation, distractor-context, jagged-frontier
- 来源: https://arxiv.org/abs/2604.15597
[2026-05-13] — ingest | ELF: Embedded Language Flows (arXiv:2605.10938, Tech Report 2026)
- 添加论文 elf-embedded-language-flows: "ELF: Embedded Language Flows" — 基于 Flow Matching 的连续嵌入语言扩散模型,用共享权重网络实现去噪-解码统一,105M 超越 170M 基线
- 新增 11 个概念页: embedded-language-flows, flow-matching, continuous-diffusion-language-models, shared-weight-discretization, classifier-free-guidance-language, self-conditioning, x-prediction-parameterization, rectified-flows, sde-sampler-language, generative-perplexity, discrete-diffusion-language-models
- 来源: https://arxiv.org/abs/2605.10938
- 作者: Keya Hu*, Linlu Qiu*, Yiyang Lu, Hanhong Zhao, Tianhong Li, Yoon Kim, Jacob Andreas, Kaiming He (MIT)
[2026-04-27] ingest | DeepSeek-V4 技术报告 (HuggingFace)
- 来源:https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro/blob/main/DeepSeek_V4.pdf
- 作者:DeepSeek-AI
- PDF:4.4MB,提取 4906 行文本
- 新增文件 (14 个):
raw/papers/deepseek-ai-deepseek-v4-2026.md— 原始论文存档papers/deepseek-v4-million-token-context.md— 论文主页面- Tier 1 核心概念 (5 个):
concepts/compressed-sparse-attention.md— CSA 压缩稀疏注意力concepts/heavily-compressed-attention.md— HCA 高强度压缩注意力concepts/manifold-constrained-hyper-connections.md— mHC 流形约束超连接concepts/muon-optimizer.md— Muon 优化器concepts/on-policy-distillation.md— OPD 在线策略蒸馏
- Tier 2 基础概念 (4 个):
concepts/hybrid-attention-architecture.md— 混合注意力架构concepts/mixture-of-experts.md— MoE 混合专家concepts/fp4-quantization-training.md— FP4 量化感知训练concepts/specialist-training-pipeline.md— 专家训练流水线
- Tier 3 占位符概念 (3 个):
concepts/multi-token-prediction.md— MTP 多 Token 预测concepts/test-time-scaling.md— 测试时扩展concepts/million-token-context.md— 百万 Token 上下文
- 关键概念:CSA/HCA 混合注意力、mHC 双随机矩阵约束、Muon 优化器、OPD 多教师蒸馏
- 更新 index.md:总页面数 57 → 71
[2026-04-20] merge | 合并 /home/ubuntu/wiki 到 /home/ubuntu/wikiplace
- 来源:旧 wiki 路径(默认回退路径 ~/wiki)
- 操作:将 wiki 独有的文件合并到 wikiplace
- 新增文件:
concepts/computerized-adaptive-testing.md— CAT 测试综述concepts/cramer-rao-lower-bound.md— CRLB 参数估计下界concepts/knowledge-bank.md— AI 辅助开发知识管理系统concepts/symbolic-regression.md— 符号回归技术raw/articles/knowledge-bank-ai-dev-2026.md— Knowledge Bank 微信公众号原文raw/papers/hbs-cramerrao-bound-notes.md— HBS CRLB 培训材料摘要raw/papers/zhuang-catsurvey-ml-2024.md— CAT 综述论文元数据raw/papers/cramerrao-bound-notes.pdf— HBS CRLB 培训 PDFraw/papers/odrzywolek-eml-universal-operator-2026.pdf— EML 论文 PDF
- 合并更新:
concepts/eml-operator.md— 补充了符号回归联系、布尔逻辑类比、研究意义和更多开放问题entities/andrzej-odrzywolek.md— 补充了发表文献、发现方法、重要意义和外部链接
- 更新 index.md:总页面数 24 → 28
- 更新 log.md:追加合并记录
[2025-04-15] create | Wiki 初始化
- 领域:数学研究、AI/ML 研究、编程技术、学习笔记与阅读资料
- 创建结构:SCHEMA.md, index.md, log.md
- 目录结构:raw/, entities/, concepts/, comparisons/, queries/
[2025-04-15] ingest | Mathematical methods and human thought in the age of AI
- 来源:arXiv:2603.26524
- 作者:terence-tao, tanya-klowden
- 保存至:raw/papers/tao-ai-mathematical-methods-2026.md
- 创建页面:
- entities/papers/tao-klowden-ai-mathematical-methods.md
- entities/terence-tao.md
- entities/tanya-klowden.md
- concepts/human-centered-ai.md
- concepts/formal-verification.md
- concepts/ai-mathematics.md
- 更新 index.md:总页面数 6
[2026-04-16] ingest | All elementary functions from a single binary operator
- 来源:arXiv:2603.21852 [cs.SC]
- 作者:andrzej-odrzywolek
- 保存至:raw/papers/odrzywolek-eml-single-operator-2026.md
- 创建页面:
- papers/odrzywolek-eml-single-operator.md — EML 算子论文摘要
- entities/andrzej-odrzywolek.md — 作者实体页面
- concepts/eml-operator.md — EML 算子概念页面
- 更新 index.md:总页面数 9
- 关键概念:EML Sheffer 算子、二叉树语法、符号回归、连续数学完备性
[2026-04-19] ingest | Memory Caching: RNNs with Growing Memory
- 来源:arXiv:2602.24281 [cs.LG]
- 作者:Ali Behrouz, Zeman Li, Yuan Deng, Peilin Zhong, Meisam Razaviyayn, Vahab Mirrokni
- 保存至:raw/papers/behrouz-memory-caching-rnn-2026.md
- 创建页面:
- papers/behrouz-memory-caching-rnn.md — MC 论文笔记
- concepts/memory-caching-rnn.md — Memory Caching 技术详解
- concepts/subquadratic-transformer-alternatives.md — 次二次 Transformer 替代方案综述
- 更新 index.md:总页面数 12
- 关键概念:Memory Caching、RNN 增长记忆、次二次复杂度、隐藏状态缓存、门控聚合
[2026-04-19] ingest | "Are You Sure?": Human Perception Vulnerability in LLM Agents
- 来源:arXiv:2602.21127 [cs.HC]
- 作者:Xinfeng Li, Shenyu Dai, Kelong Zheng, Yue Xiao, Gelei Deng, Wei Dong, Xiaofeng Wang
- 保存至:raw/papers/li-amd-human-perception-2026.md
- 创建页面:
- papers/li-amd-human-perception.md — AMD 实证研究论文笔记
- concepts/agent-mediated-deception.md — AMD 攻击模式详解
- concepts/human-agent-trust.md — 人机信任与脆弱性
- 更新 index.md:总页面数 14
- 关键概念:Agent-Mediated Deception、HAT-Lab、认知失败模式、经验学习、信任校准
[2026-04-19] ingest | Prefill-as-a-Service: KVCache Goes Cross-Datacenter
- 来源:arXiv:2604.15039 [cs.DC]
- 作者:Ruoyu Qin, Weiran He, Yaoyu Wang, Zheming Li, Xinran Xu, Yongwei Wu, Weimin Zheng, Mingxing Zhang
- 保存至:raw/papers/qin-prfaas-cross-datacenter-2026.md
- 创建页面:
- papers/qin-prfaas-cross-datacenter.md — PrfaaS 论文笔记
- concepts/prefill-as-a-service.md — PrfaaS 架构详解
- concepts/prefill-decode-disaggregation.md — PD 分离架构演进
- concepts/kvcache-transfer.md — KVCache 传输与优化
- 更新 index.md:总页面数 17
- 关键概念:Prefill-as-a-Service、跨数据中心部署、KVCache 传输、混合注意力、带宽感知调度
[2026-04-19] ingest | Mixture-of-Depths Attention (MoDA)
- 来源:arXiv:2603.15619 [cs.LG]
- 作者:Lianghui Zhu, Yuxin Fang, Bencheng Liao, Shijie Wang, Tianheng Cheng, Zilong Huang, Chen Chen, Lai Wei, Yutao Zeng, Ya Wang, Yi Lin, Yu Li, Xinggang Wang
- 保存至:raw/papers/zhu-moda-mixture-of-depths-2026.md
- 创建页面:
- papers/zhu-moda-mixture-of-depths.md — MoDA 论文笔记
- concepts/mixture-of-depths-attention.md — MoDA 机制详解
- concepts/depth-scaling-signal-degradation.md — 深度扩展与信号退化问题
- 更新 index.md:总页面数 21
- 关键概念:Mixture-of-Depths Attention、信号退化、跨层 KV 访问、硬件高效实现、Post-Norm 优势
[2026-04-19] ingest | OPPO 多模态数据湖实践 (WeChat Article)
- 来源:微信公众号文章 (DataFun / Data for AI Meetup)
- 分享人:David (OPPO 大数据架构负责人)
- 链接:https://mp.weixin.qq.com/s/cBaYa04qAIGsxG1hD7ll3w
- 保存至:raw/articles/oppo-multimodal-data-lake-2026.md
- 创建页面:
- articles/oppo-multimodal-data-lake.md — 文章核心架构与成果总结
- concepts/gravitino-unified-metadata.md — Gravitino 统一元数据管理
- concepts/curvine-distributed-cache.md — Curvine 分布式缓存系统
- 更新 index.md:新增 Articles 分区,总页面数 24
- 关键概念:多模态数据湖、Gravitino 元数据、Curvine 缓存、LanceDB 加速、混合云架构
[2026-04-20] ingest | Spurious Predictability in Financial Machine Learning
- 来源:arXiv:2604.15531 [q-fin.ST, stat.ME, stat.ML]
- 作者:Sotirios D. Nikolopoulos
- 保存至:raw/papers/nikolopoulos-spurious-predictability-2026.md
- 创建页面:
- papers/nikolopoulos-spurious-predictability.md — 金融机器学习虚假可预测性论文笔记
- concepts/spurious-predictability.md — 虚假可预测性概念详解
- 更新 index.md:总页面数 30
[2026-04-20] ingest | Hyperagents: Self-Referential Agents with Metacognitive Self-Modification
- 来源:arXiv:2603.19461 [cs.AI]
- 作者:Jenny Zhang, Bingchen Zhao, Wannan Yang, Jakob Foerster, Jeff Clune, Minqi Jiang, Sam Devlin, Tatiana Shavrina
- 保存至:raw/papers/zhang-hyperagents-2026.md
- 创建页面:
- papers/zhang-hyperagents.md — 超智能体论文笔记
- concepts/hyperagents.md — 超智能体概念详解
- concepts/self-improving-ai.md — 自我改进人工智能概念
- concepts/darwin-godel-machine.md — 达尔文·哥德尔机概念
- concepts/metacognitive-self-modification.md — 元认知自我修改概念
- 更新 index.md:总页面数 35
- 关键概念:超智能体、自我改进 AI、达尔文·哥德尔机、元认知自我修改、自我加速进展、可编辑元级
[2026-04-20] fix | 修复超智能体相关概念的断链
- 修复问题:新创建页面中存在指向未创建概念的链接
- 创建缺失概念页面:
- concepts/meta-learning.md — 元学习概念
- concepts/recursive-self-improvement.md — 递归自我改进概念
- concepts/genetic-programming.md — 遗传编程概念
- concepts/program-synthesis.md — 程序合成概念
- concepts/cognitive-architecture.md — 认知架构概念
- concepts/singularity.md — 技术奇点概念
- 创建占位符概念页面(修复剩余断链):
- concepts/ai-alignment.md — AI 对齐概念
- concepts/ai-safety.md — AI 安全概念
- concepts/neuroscience.md — 神经科学概念
- concepts/evolutionary-algorithms.md — 进化算法概念
- concepts/few-shot-learning.md — 少样本学习概念
- concepts/transfer-learning.md — 迁移学习概念
- 更新 index.md:总页面数 46
- 修复效果:消除所有新页面中的断链,建立完整的概念网络
- 关键概念:虚假可预测性、证伪审计、选择诱导性能膨胀、有效多重性、金融机器学习方法论
[2026-04-22] ingest | ClawLess: A Security Model of AI Agents
- 来源:arXiv:2604.06284v1 [cs.CR]
- 作者:Hongyi Lu, Nian Liu, Shuai Wang, Fengwei Zhang
- 机构:南方科技大学,香港科技大学
- 保存至:raw/papers/lu-hongyi-clawless-ai-agent-security-2026.md
- 创建页面:
- papers/clawless-ai-agent-security.md — ClawLess 论文笔记
- concepts/clawless.md — ClawLess 安全框架概念
- concepts/ai-agent-security.md — AI 代理安全概念
- concepts/formal-security-model.md — 形式化安全模型概念
- concepts/userspace-kernel.md — 用户空间内核概念
- concepts/bpf-syscall-interception.md — BPF系统调用拦截概念
- concepts/secure-containers.md — 安全容器概念
- concepts/worst-case-threat-model.md — 最坏情况威胁模型概念
- 更新 index.md:总页面数 46 → 53
- 关键概念:ClawLess、AI代理安全、形式化安全模型、用户空间内核、BPF系统调用拦截、安全容器、最坏情况威胁模型
[2026-04-22] ingest | Crawl4AI: 开源智能网页爬虫与数据提取工具
- 来源:知乎专栃 https://zhuanlan.zhihu.com/p/717965307
- 作者:沈飞
- 保存至:raw/articles/shenfei-crawl4ai-open-source-web-crawler-2024.md
- 创建页面:
- articles/crawl4ai-open-source-web-crawler.md — Crawl4AI 文章主页面
- concepts/crawl4ai.md — Crawl4AI 工具概念页面
- concepts/rag-systems.md — RAG 系统概念页面
- concepts/llm-applications.md — LLM 应用概念页面
- 更新 index.md:总页面数 53 → 57
- 关键概念:Crawl4AI、网页爬虫、数据提取、RAG、LLM应用、Markdown转换
2026-04-28 | 哥德尔不完备定理教程
- 来源: PDF 直接提交 (godel_tutorial.pdf),2026年4月综合教程
- 作者: 无明确单一作者(面向数学系本科生的教学资料)
- 新增页面: 25 个(1 论文 + 1 原始存档 + 23 概念)
- raw/papers/godel-tutorial-2026.md — 原始存档
- papers/godel-incompleteness-tutorial.md — 论文主页面
- concepts/godel-incompleteness-theorems.md — 哥德尔不完备定理
- concepts/godel-numbering.md — 哥德尔编码
- concepts/hilberts-program.md — 希尔伯特计划
- concepts/peano-arithmetic.md — 皮亚诺算术
- concepts/self-reference.md — 自指
- concepts/diagonalization-method.md — 对角线方法
- concepts/halting-problem.md — 停机问题
- concepts/lucas-penrose-argument.md — 卢卡斯-彭罗斯论证
- concepts/chaitin-algorithmic-information-theory.md — 算法信息论
- concepts/metamathematics.md — 元数学
- concepts/primitive-recursive-functions.md — 原始递归函数
- concepts/computability-theory.md — 可计算性理论
- concepts/formal-systems.md — 形式系统
- concepts/automated-theorem-proving.md — 自动定理证明
- concepts/paris-harrington-theorem.md — 巴黎-哈灵顿定理
- concepts/goodsteins-theorem.md — 古德斯坦定理
- concepts/russells-paradox.md — 罗素悖论
- concepts/continuum-hypothesis.md — 连续统假设
- concepts/consistency-logic.md — 一致性
- concepts/completeness-logic.md — 完备性
- concepts/mathematical-pluralism.md — 数学多元主义
- concepts/chaitin-constant.md — 蔡廷常数
- concepts/kolmogorov-complexity.md — 柯尔莫哥洛夫复杂度
- 更新 index.md:总页面数 71 → 96
- 关键概念:哥德尔不完备定理、哥德尔编码、自指、对角线方法、停机问题、希尔伯特计划、可计算性、形式系统
[2026-04-29] ingest | 大语言模型注意力机制全面分析 (综述论文)
- 来源:用户直接上传 PDF (LLM注意力机制全面分析.pdf)
- 类型:综述论文 / Review Paper,2026年4月
- PDF:1385 行文本提取
- 新增文件 (21 个):
raw/papers/llm-attention-survey-2026.md— 原始论文存档papers/llm-attention-survey-2026.md— 论文主页面- Tier 1 核心概念 (6 个):
concepts/multi-head-attention.md— MHA 标准多头注意力concepts/grouped-query-attention.md— GQA 分组查询注意力concepts/multi-head-latent-attention.md— MLA 多潜在头注意力concepts/flash-attention.md— FlashAttention IO感知优化concepts/attention-entropy-collapse.md— 注意力熵崩溃concepts/kv-cache-bottleneck.md— KV缓存内存瓶颈
- Tier 2 基础概念 (5 个):
concepts/multi-query-attention.md— MQA 多查询注意力concepts/sparse-attention-patterns.md— 稀疏注意力模式concepts/linear-attention-methods.md— 线性注意力方法concepts/rotary-position-embedding.md— RoPE 旋转位置编码concepts/lost-in-the-middle.md— Lost in the Middle 现象
- Tier 3 占位概念 (8 个):
concepts/attention-sinks.md— 注意力汇concepts/flash-attention-3.md— FlashAttention-3concepts/mamba-ssm.md— Mamba 状态空间模型concepts/mixture-of-attention-schemes.md— MoAS 注意力方案混合concepts/duo-attention.md— DuoAttention 双模式注意力concepts/seer-attention.md— SeerAttention 可学习稀疏concepts/ntk-aware-interpolation.md— NTK-aware 位置插值concepts/native-sparse-attention.md— NSA 原生稀疏注意力
- 更新 index.md:总页面数 96 → 116
- 关键概念:注意力机制演化谱系 (MHA→MQA→GQA→MLA)、FlashAttention、注意力退化、KV缓存瓶颈、Lost in the Middle
- 网络连接:与已有概念 CSA、HCA、混合注意力架构、DeepSeek-V4 等形成密集交叉引用
[2026-04-29] ingest | GPT-Image-2 绘图 Prompt 方法论与风格合集
- 来源:linux.do 论坛 (sallyn),https://linux.do/t/topic/2044964
- 类型:论坛教程/经验分享 (2026-04-24),整理于 2026-04-28
- 新增文件 (11 个):
raw/articles/sallyn-gpt-image2-prompt-collection-2026.md— 原始摘录存档articles/gpt-image2-prompt-collection.md— 文章主页面- Tier 1 核心概念 (3 个):
concepts/gpt-image2.md— GPT-Image-2 图像生成工具concepts/prompt-reverse-engineering.md— 图片反推 Prompt:15维分析框架concepts/image-generation-prompt-design.md— 图像生成 Prompt 设计方法论
- Tier 2 风格概念 (6 个):
concepts/russian-constructivism.md— 俄国构成主义concepts/glitch-art-style.md— 故障艺术concepts/cel-shading-style.md— 赛璐璐风格concepts/risograph-print-style.md— Riso印刷风格concepts/halftone-print-style.md— 半调印刷风格concepts/klein-blue.md— 克莱因蓝
- 更新 index.md:总页面数 116 → 126
- 关键概念:GPT-Image-2、Prompt反推工程、15维美学分析框架、5种核心艺术风格
- 特色:首次将 AI 图像生成工具链和艺术风格概念纳入 wiki 知识网络
[2026-04-29] ingest | Caddy 反向代理认证方案
- 来源:用户直接上传 TXT
- 类型:技术教程/配置指南
- 新增文件 (6 个):
raw/articles/caddy-reverse-proxy-auth-2026.md— 原始文档存档articles/caddy-reverse-proxy-auth.md— 文章主页面- 概念 (4 个):
concepts/caddy-web-server.md— Caddy Web 服务器concepts/reverse-proxy-authentication.md— 反向代理层认证模式concepts/api-key-authentication.md— API Key 认证机制concepts/forward-authentication.md— 外部委托认证模式
- 更新 index.md:总页面数 126 → 131
- 关键概念:命名匹配器、反向代理认证、API Key 白名单、forward_auth 委托
- 特色:首次将 Web 服务器/反向代理/认证基础设施概念纳入 wiki
[2026-04-29] ingest | How Far Can Unsupervised RLVR Scale LLM Training? (arXiv:2603.08660)
- 来源:arXiv API (2603.08660)
- 作者:He, Zuo, Liu et al. (22 authors, Tsinghua/Shanghai AI Lab et al.)
- 会议:ICLR 2026
- PDF:7121 行文本提取
- 新增文件 (13 个):
raw/papers/he-urlvr-sharpening-2026.md— 原始存档papers/he-urlvr-sharpening-2026.md— 论文主页面- Tier 1 核心概念 (4 个):
concepts/unsupervised-rlvr.md— URLVR 范式定义concepts/intrinsic-rewards-sharpening.md— Sharpening 统一理论concepts/model-collapse-step.md— MCS 模型崩溃步concepts/self-verification-rewards.md— 自我验证外部奖励
- Tier 2 基础概念 (4 个):
concepts/reward-hacking-llm.md— 奖励黑客与模型崩溃concepts/certainty-based-rewards.md— 确定性奖励concepts/ensemble-based-rewards.md— 集成奖励/多数投票concepts/generation-verification-asymmetry.md— 生成-验证不对称性
- Tier 3 占位概念 (3 个):
concepts/rlvr-unified-framework.md— RLVR 统一框架concepts/test-time-training-rl.md— 测试时训练 RLconcepts/confidence-correctness-alignment.md— 置信度-正确性对齐
- 更新 index.md:总页面数 131 → 143
- 关键概念:URLVR、Sharpening机制、Rise-then-Fall模式、Model Collapse Step、Self-verification突破
- 特色:首次将 RLVR/URLVR/奖励黑客等 LLM 后训练理论概念纳入 wiki
2026-04-30 20:08 — Thinking with Visual Primitives (DeepSeek-AI, 2026)
来源: GitHub (deepseek-ai/Thinking-with-Visual-Primitives) 类型: 技术报告 / 研究论文 领域: Multimodal AI / Visual Reasoning
新增页面 (21)
- Papers: thinking-with-visual-primitives — 视觉原语思考框架主页面
- Raw: raw/papers/deepseek-visual-primitives-2026.md
新增概念 (20)
- visual-primitives — 视觉原语:框+点作为思维最小单位
- reference-gap — 引用鸿沟:语言空间指代模糊
- perception-gap — 感知鸿沟:分辨率限制的视觉细节丢失
- chain-of-thought — 思维链 (CoT) 的多模态扩展
- multimodal-large-language-model — MLLM 背景概念
- system-2-thinking — System 2 思维与视觉推理
- deepseek-vit — DeepSeek 视觉 Transformer
- deepseek-v4-flash — 语言骨干模型
- token-efficiency — Token 效率 (7056× 压缩)
- coarse-grained-counting — 粗粒度计数
- fine-grained-counting — 细粒度计数
- maze-navigation — 迷宫导航
- path-tracing — 路径追踪
- group-relative-policy-optimization — GRPO 算法
- specialized-sft — 专项监督微调
- specialized-rl — 专项强化学习
- unified-rft — 统一拒绝采样微调
- exponential-decay-reward — 指数衰减奖励
- bidirectional-trajectory-evaluation — 双向轨迹评估
- reward-model — 奖励模型体系
交叉链接
与已有概念 compressed-sparse-attention、on-policy-distillation、mixture-of-experts、deepseek-v4-million-token-context 建立双向链接。
Wiki 规模
143 → 164 页
[2026-05-01] ingest | CL-Bench Life: 真实生活上下文学习基准
- 来源:arXiv:2604.27043 [cs.CL]
- 作者:Hunyuan Team (Tencent) & Fudan University
- 日期:2026-04-29
- PDF:4.9MB,提取 3879 行文本
- 新增文件 (10 个):
raw/papers/hunyuan-team-cl-bench-life-2026.md— 原始论文存档papers/hunyuan-team-cl-bench-life.md— 论文主页面- Tier 1 核心概念 (3 个):
concepts/cl-bench-life.md— CL-bench Life 基准设计concepts/real-life-context-learning.md— 真实生活上下文学习能力concepts/context-misuse.md— 上下文误用:首要失败模式
- Tier 2 基础概念 (1 个):
concepts/messy-context-reasoning.md— 混乱上下文推理
- Tier 2/3 占位概念 (4 个):
concepts/context-learning.md— 通用上下文学习concepts/llm-evaluation-benchmarks.md— LLM 评测基准体系concepts/long-context-understanding.md— 长上下文理解concepts/identity-reference-resolution.md— 身份指代消解
- 更新 index.md:总页面数 164 → 173
- 关键概念:真实生活上下文学习、CL-bench Life、上下文误用(76-84%错误)、混乱上下文推理、三大上下文类别
- 核心发现:最佳模型仅 19.3% 解决率;上下文误用是首要失败模式;长上下文能力与混乱上下文推理不等价
[2026-05-01] lint | Wiki 全面健康检查与大修
- 检查范围:181 个 wiki 页面
- 修复前问题:462 total(117 断链 + 121 索引重复 + 106 缺失 frontmatter + 18 孤儿 + 等)
- 修复操作:
- 索引去重:732 条概念条目 → 154 条唯一,26 条论文 → 15 条唯一,文件从 810 行压缩到 198 行
- 断链清零:117 → 0,批量修复中文 wikilink 目标错误(ClawLess 系列、Tao/Klowden 系列等)
- 缺失索引条目:补回 5 个概念 + 4 篇文章 + 清理 2 个坏条目
- Frontmatter 补全:106 → 0,全量补充 YAML frontmatter
- 孤儿概念链接:3 个 URLVR 相关概念加回 inbound link
- 文件移动:entities/papers/tao-klowden-ai-mathematical-methods.md → papers/
- 修复后状态:
- 断链:0 ✅
- 缺失 frontmatter:0 ✅
- 索引条目:173,声明总数:181(差值 8 为 reviews/extracts)
- 孤儿:7(全部为 reviews/extracts,有意设计)
- 页面数不变:181
[2026-05-01] ingest | Agent网络三层分类法综述
- 来源:TechRxiv (DOI: 10.36227/techrxiv.177127384.46731320/v1)
- 作者:Xinyuan Song (Emory), Qingsong Wen (Oxford), Shirui Pan (Griffith), Liang Zhao (Emory)
- 日期:2026-02-16
- PDF:用户直接上传,提取 2084 行文本
- 新增文件 (9 个):
raw/papers/song-agent-network-taxonomy-2026.md— 原始论文存档papers/song-agent-network-taxonomy.md— 论文主页面- Tier 1 核心概念 (4 个):
concepts/agent-network-taxonomy.md— 三层级分类法concepts/agent-network-topology.md— 拓扑维度(集中式vs去中心化)concepts/agent-network-memory-scope.md— 记忆范围维度(全局vs局部)concepts/agent-network-update-behavior.md— 更新行为维度(静态vs动态)
- Tier 2 基础概念 (3 个):
concepts/centralized-agent-architecture.md— 集中式架构详解concepts/decentralized-agent-architecture.md— 去中心化架构详解concepts/agent-communication-stack.md— 三层通信协议栈
- 交叉链接:与 cognitive-architecture、hyperagents 建立双向链接
- 更新 index.md:总页面数 181 → 189
- 关键概念:Agent网络三层分类法、8种系统类别、通信协议栈、MCP标准化
- 核心贡献:嵌套式分类框架(A=(V,E,M,Π))→8种类别;识别语义层为大规模系统首要失败点
[2026-05-01] ingest | CL-bench: 首个上下文学习基准
- 来源:arXiv:2602.03587 [cs.CL]
- 作者:Shihan Dou, Ming Zhang, Zhangyue Yin et al. (27 authors, Fudan Univ. & Tencent Hunyuan)
- 日期:2026-02-03
- PDF:1.8MB,提取 6713 行文本
- 新增/更新文件 (7 个):
raw/papers/dou-cl-bench-2026.md— 原始论文存档papers/dou-cl-bench.md— 论文主页面concepts/context-learning.md— 从占位页升级为完整概念页- Tier 1 类别概念 (4 个):
concepts/domain-knowledge-reasoning.md— 领域知识推理(7子类)concepts/rule-system-application.md— 规则系统应用(5子类)concepts/procedural-task-execution.md— 程序性任务执行(3子类)concepts/empirical-discovery-simulation.md— 经验发现与模拟(3子类)
- 更新 index.md:总页面数 189 → 195
- 关键概念:Context Learning 范式定义、CL-bench 四类别框架、污染防护设计
- 核心发现:十模型平均 17.2%/最佳 23.7%;归纳推理(经验发现)是最瓶颈;法律推理 >40% vs 数学形式化 <15%
- 与已有概念的连接:与 cl-bench-life、real-life-context-learning、context-misuse 形成 CL-bench 系列完整知识网络
[2026-05-11] ingest | Prompt Caching 架构工程手册 (微信公众号)
- 来源:https://mp.weixin.qq.com/s/gyd4cqxadv3YW5Fe09r95g
- 类型:工程实践教程 (Article)
- 案例系统:Meta-JCTrader(高频交易 + RL + Meta-Learning)
- 新增文件 (15 个):
raw/articles/prompt-caching-architecture-2026.md— 原始文章存档articles/prompt-caching-architecture.md— 文章主页面- 核心概念 (12 个):
concepts/prompt-caching.md— Prompt Cachingconcepts/prefix-matching.md— 前缀匹配concepts/prompt-layering.md— 提示分层 (Global/Project/Session/Dynamic)concepts/stub-pattern.md— Stub 模式(轻量化桩)concepts/tool-registry.md— ToolRegistry 统一接口concepts/cache-safe-forking.md— 缓存安全分叉concepts/cache-invalidation.md— 缓存失效concepts/cache-hit-ratio.md— 缓存命中率 (CHR)concepts/context-compression.md— 上下文压缩concepts/system-message-abuse.md— System Message 滥用反模式concepts/cache-health-observability.md— 缓存健康度可观测性concepts/meta-jctrader.md— Meta-JCTrader 案例
- 占位符概念 (2 个):
concepts/agentic-systems.md— Agentic Systemsconcepts/reinforcement-learning-trading.md— 强化学习交易
- 索引:195 → 203 页(全量重建)
- 关键概念:四层架构分层、Stub模式/ToolRegistry、Cache-Safe Forking、CHR监控
- Review:
reviews/prompt-caching-architecture-review-20260511.md
[2026-05-11] ingest | 拉姆齐数的数学综述 (用户上传)
- 来源:用户上传 Markdown (RNS.md)
- 日期:2025年6月
- 类型:数学综述 (Survey)
- 新增文件 (18 个):
raw/papers/ramsey-numbers-survey-2025.md— 原始综述存档papers/ramsey-numbers-survey.md— 论文主页面- 核心概念 (12 个):
concepts/ramsey-theory.md— 拉姆齐理论concepts/ramsey-numbers.md— 拉姆齐数concepts/diagonal-ramsey-number.md— 对角拉姆齐数concepts/probabilistic-method.md— 概率方法 (Erdős 1947)concepts/hypergraph-ramsey-number.md— 超图拉姆齐数concepts/geometric-ramsey-theory.md— 几何拉姆齐理论concepts/additive-combinatorics.md— 加法组合学concepts/van-der-waerden-theorem.md— van der Waerden 定理concepts/paris-harrington-theorem.md— 巴黎-哈灵顿定理concepts/green-tao-theorem.md— Green-Tao 定理 (素数等差数列)concepts/szemerédi-regularity-lemma.md— Szemerédi 正则性引理concepts/ramsey-theory-applications.md— 拉姆齐理论跨学科应用
- 占位符概念 (4 个):
concepts/paley-graph.md— Paley 图concepts/lovasz-local-lemma.md— Lovász 局部引理concepts/random-graph-theory.md— 随机图理论concepts/furstenberg-correspondence.md— Furstenberg 对应原理
- 索引:203 → 219 页(全量重建)
- 关键概念:Ramsey 理论核心信条、概率方法、Green-Tao 定理、Paris-Harrington 不可判定性
- Review:
reviews/ramsey-numbers-survey-review-20260511.md - 与已有概念的连接:godel-incompleteness-theorems (via Paris-Harrington)
[2026-05-11] ingest | 上下文构造与拉姆齐数 (用户上传)
- 来源:用户上传 Markdown
- 类型:方法论设计 (Methodology)
- 核心思路:将拉姆齐理论的"必然涌现的秩序"映射到 Agent 上下文构筑
- 新增文件 (7 个):
raw/articles/ramsey-context-construction-2026.md— 原始文档存档articles/ramsey-context-construction.md— 方法论主页面- 核心概念 (5 个):
concepts/ramsey-context-graph.md— 拉姆齐上下文图(蓝/红边兼容性建模)concepts/ramsey-context-cache.md— 拉姆齐上下文缓存(三层机制)concepts/context-blue-clique.md— 上下文蓝色团(全兼容骨架)concepts/greedy-context-screening.md— 贪心上下文筛选(三步快速组装)concepts/ramsey-context-template.md— 拉姆齐上下文模板(KV cache 优化)
- 索引:219 → 225 页(全量重建)
- 关键概念:兼容图建模、R(3,3)=6 保证、蓝色团模板、贪心团搜索
- Review:
reviews/ramsey-context-construction-review-20260511.md - 桥梁作用:连接 ramsey-theory(数学)与 prompt-caching(工程)
[2026-05-11] ingest | Koopa: Koopman 预测器驱动的非平稳时序学习 (arXiv)
- 来源:https://arxiv.org/abs/2305.18803
- 作者:Yong Liu, Chenyu Li, Jianmin Wang, Mingsheng Long (Tsinghua)
- 会议:NeurIPS 2023
- 新增文件 (9 个):
raw/papers/liu-koopa-2023.md— 原始论文存档papers/liu-koopa-2023.md— 论文主页面- 核心概念 (7 个):
concepts/koopman-theory.md— Koopman 理论(非线性→线性映射)concepts/koopman-predictor.md— Koopman 预测器concepts/fourier-filter-dynamics.md— Fourier Filter 动力学分解concepts/dynamic-mode-decomposition.md— DMD 动态模式分解concepts/non-stationary-time-series.md— 非平稳时间序列concepts/koopman-autoencoder.md— Koopman 自编码器 (KAE)concepts/time-variant-dynamics.md— 时变动力学
- 索引:225 → 233 页(全量重建)
- 关键结果:SOTA 竞争性能 + 77.3% 训练时间节省 + 76.0% 内存节省
- Review:
reviews/koopa-review-20260511.md