# AI Research Radar - 2026-08-17 - Profile: George Research Profile v2 - Summary mode: single - Provider: deepseek - Model: deepseek-v4-flash - LLM summary calls: 7 - Estimated cost: RMB 0.0 / 1.0 - Last LLM error: provider=deepseek; model=deepseek-v4-flash; base_url=https://api.deepseek.com; HTTP status=n/a; error=Could not parse JSON response: - provider_disabled: kimi - reason: unauthorized ## 0. Daily Overview - Most important direction: 上下文压缩 / 长上下文 / 记忆 - Must Read count: 0 - Skim count: 8 (Teaching LLMs to Update Beliefs for Efficient Long-Horizon Interaction; Adaptive Parallel Reasoning: The Next Paradigm in Efficient Inference Scaling; LycheeMemory V2: Efficient Long-Term Memory for LLM Agents via Semantic Segment-Level Consolidation; Mind the Context: Continual Learning of Socially Appropriate Robot Actions via Environmental-Social Disentanglement; Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillation) - Watch count: 12 (2026 BAIR Graduate Showcase; Identifying Interactions at Scale for LLMs; Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence; Vero: Can AI Agents Build Formally Verified Software Repositories?; CoverPrune: Coverage-Driven Token Pruning for 3D VLMs via Optimal Transport) - Keywords: agent, framework, reasoning, agent memory, nlp, robotics, attention, berkeley.edu - Judgement: 今日主线: 没有强制深读项, 建议归档观察. ## 1. Core Research Tracks ### 1.1 AI Systems / HPC / Distributed Training & Inference #### Must Read - 无。 #### Skim - 无。 #### Watch - [Identifying Interactions at Scale for LLMs](http://bair.berkeley.edu/blog/2026/03/13/spex/) （关注；AI 系统 / HPC / 分布式训练与推理；个人相关度=0.87；全局热度=0.41；炒作风险=0.00） - [When Local Variance Optimality Is Not Enough: RoPE-Aligned Q/K Rotations for Dynamic 4-Bit Quantisation](https://arxiv.org/abs/2608.13365v1) （关注；AI 系统 / HPC / 分布式训练与推理；个人相关度=0.83；全局热度=0.43；炒作风险=0.00） - [Intern-S2-Preview: Scientific Agentic Foundation Model](https://arxiv.org/abs/2608.13505v1) （关注；AI 系统 / HPC / 分布式训练与推理；个人相关度=0.82；全局热度=0.37；炒作风险=0.00） ### 1.2 GPU-Centric I/O / Networking / Storage #### Must Read - 无。 #### Skim - 无。 #### Watch - [The Ingestion Tax: Adopting File-Backed Weights in Tensor Frameworks](https://arxiv.org/abs/2608.12114v1) （关注；GPU 中心 I/O / 网络 / 存储；个人相关度=0.81；全局热度=0.35；炒作风险=0.00） - [TopoIntent: Compiling Security Intent into Executable, Compliance-Checked Network Topologies](https://arxiv.org/abs/2608.13389v1) （关注；GPU 中心 I/O / 网络 / 存储；个人相关度=0.79；全局热度=0.36；炒作风险=0.00） ### 1.3 Compression / Reliability for AI Infrastructure #### Must Read - 无。 #### Skim - 无。 #### Watch - [FaCTz: Fast Critical-Point and Topology-Aware GPU Compression for Scientific Vector Fields](https://arxiv.org/abs/2608.10586v1) （关注；AI 基础设施压缩 / 可靠性；个人相关度=0.81；全局热度=0.34；炒作风险=0.00） - [Exponential quantum advantage for learning signals with a single qubit](https://arxiv.org/abs/2608.13521v1) （关注；AI 基础设施压缩 / 可靠性；个人相关度=0.80；全局热度=0.36；炒作风险=0.00） - [The data geometry of masking diffusion: Certified-optimal schedules via unmasking growth complexity](https://arxiv.org/abs/2608.13520v1) （关注；AI 基础设施压缩 / 可靠性；个人相关度=0.77；全局热度=0.35；炒作风险=0.00） ### 1.4 Agent Runtime / RL Infrastructure / Scheduling #### Must Read - 无。 #### Skim - 无。 #### Watch - [Vero: Can AI Agents Build Formally Verified Software Repositories?](https://arxiv.org/abs/2608.13522v1) （关注；Agent 运行时 / RL 基础设施 / 调度；个人相关度=0.85；全局热度=0.38；炒作风险=0.00） - [MARC v1: An Open-Source Multi-Agent Framework for Clinical AI Reasoning and Coordination](https://arxiv.org/abs/2608.13476v1) （关注；Agent 运行时 / RL 基础设施 / 调度；个人相关度=0.80；全局热度=0.38；炒作风险=0.00） - [Training AI Scientists to Replicate Research](https://arxiv.org/abs/2608.13331v1) （关注；Agent 运行时 / RL 基础设施 / 调度；个人相关度=0.78；全局热度=0.34；炒作风险=0.00） ### 1.5 Embodied Intelligence / VLA / World Models #### Must Read - 无。 #### Skim ##### 1. [Mind the Context: Continual Learning of Socially Appropriate Robot Actions via Environmental-Social Disentanglement](https://arxiv.org/abs/2608.13448v1) - Reading tier: SKIM - Source: arXiv AI/ML/NLP/Vision/Robotics (primary; role=paper_source) - Published: 2026-08-13T16:33:10+00:00 - Primary track: Embodied Intelligence / VLA / World Models - Secondary tags: 其他亮点, Novel Class Discovery / Open-World Learning / OOD / Continual Learning, AI Systems / HPC / Distributed Training & Inference, Agent Runtime / RL Infrastructure / Scheduling - Grounding level: abstract only - Scores: personal=0.81, global=0.47, credibility=1.00, evidence=1.00, hype_risk=0.00, feedback=0.00 - Project relevance: skyfs=0.00, schedagent=0.00, verl_infrastructure=0.00, embodied_intelligence=0.22 - What it is: Mind the Context: Continual Learning of Socially Appropriate Robot Actions via Environmental-Social Disentanglement: 研究论文, 方向为“Embodied Intelligence / VLA / World Models”; 主要线索: agent, continual learning, cs.RO, framework. - Problem: 它关注“Embodied Intelligence / VLA / World Models”里的 agent, continual learning, cs.RO, framework 等问题. - Method/contribution: 摘要可确认它偏向评测或数据构建; 具体任务定义, 指标和样本规模需读原文确认. - Why important to George: Reading tier: SKIM editorial_priority: 0.74 今天快速扫读. personal: 0.81, relevance: 1.00. - Suggested action: skim - Matched keywords: agent, continual learning, cs.RO, framework, github, nlp, robot, robotics #### Watch - [2026 BAIR Graduate Showcase](http://bair.berkeley.edu/blog/2026/07/01/grads-2026/) （关注；具身智能 / VLA / 世界模型；个人相关度=0.97；全局热度=0.41；炒作风险=0.00） - [CoverPrune: Coverage-Driven Token Pruning for 3D VLMs via Optimal Transport](https://arxiv.org/abs/2608.13226v1) （关注；具身智能 / VLA / 世界模型；个人相关度=0.83；全局热度=0.36；炒作风险=0.00） - [AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design](https://arxiv.org/abs/2608.13560v1) （关注；具身智能 / VLA / 世界模型；个人相关度=0.83；全局热度=0.38；炒作风险=0.00） ## 2. Supporting AI Foundations ### Context / Memory - [Full-bandwidth transformer](https://arxiv.org/abs/2608.08888) （关注；上下文压缩 / 长上下文 / 记忆；个人相关度=0.67；全局热度=0.43；炒作风险=0.00） ### Generic Agents / Reasoning - [Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence](https://arxiv.org/abs/2608.12743) （关注；Agent / 推理 / 推理时扩展 / 规划；个人相关度=0.86；全局热度=0.49；炒作风险=0.00） - [AVA-Encoder: Towards Agent-Native Video Representation Learning](https://arxiv.org/abs/2608.12313) （关注；Agent / 推理 / 推理时扩展 / 规划；个人相关度=0.82；全局热度=0.50；炒作风险=0.00） - [AaLLM: An End-to-End Analog Circuit Design Framework from Topology Generation to Sizing Using Large Language Models](https://arxiv.org/abs/2608.13472v1) （关注；Agent / 推理 / 推理时扩展 / 规划；个人相关度=0.81；全局热度=0.37；炒作风险=0.00） ### Reinforcement Learning - [Parameter Exploration for RLVR via Variational Learning](https://arxiv.org/abs/2608.09805) （归档；RL；个人相关度=0.59；全局热度=0.45；炒作风险=0.00） - [Import AI 460: Reward hacking society, RSI data from Anthropic; and RL-based quadcopter racing](https://jack-clark.net/2026/06/08/import-ai-460-reward-hacking-society-rsi-data-from-anthropic-and-rl-based-quadcopter-racing/) （归档；RL；个人相关度=0.38；全局热度=0.30；炒作风险=0.28） ### Model Architecture - [Maglev: Sliding Recurrent Memory](https://arxiv.org/abs/2608.02870) （归档；模型架构；个人相关度=0.60；全局热度=0.42；炒作风险=0.00） - [Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus](https://arxiv.org/abs/2608.12149) （归档；模型架构；个人相关度=0.57；全局热度=0.48；炒作风险=0.00） ### Multimodal / VLM / CV - [AmalthAI: An Open-Source Computer Vision Platform for Cultural Heritage](https://arxiv.org/abs/2608.13343v1) （关注；CV；个人相关度=0.73；全局热度=0.37；炒作风险=0.00） - [LiveAnimate: Stable Long-Form Streaming Human Animation in Real-Time](https://arxiv.org/abs/2608.11745) （关注；CV；个人相关度=0.69；全局热度=0.49；炒作风险=0.00） ### NLP - [Motor, Cognitive, or Corpus? What Survives Cross-Lingual Transfer in Speech-Based Parkinsons Disease Detection](https://arxiv.org/abs/2608.13425v1) （归档；NLP；个人相关度=0.64；全局热度=0.36；炒作风险=0.00） - [CROP: Task Relevance via Counterfactuals for Selective On-Policy Distillation](https://arxiv.org/abs/2608.13387v1) （归档；NLP；个人相关度=0.58；全局热度=0.34；炒作风险=0.00） ### Open-World / Continual Learning - 无。 ### Model Distillation - [PixSDS: Why Latent SDS Makes Noisy Pixels](https://arxiv.org/abs/2608.12997) （关注；模型蒸馏 / 模型压缩；个人相关度=0.69；全局热度=0.47；炒作风险=0.00） - [Locking Pretrained Weights via Deep Low-Rank Residual Distillation](https://machinelearning.apple.com/research/locking-pretrained-weights) （关注；模型蒸馏 / 模型压缩；个人相关度=0.66；全局热度=0.29；炒作风险=0.00） ## 3. Cross-Track Connections - VLA inference latency ↔ GPU serving - robot rollout ↔ RL infrastructure - world model simulation ↔ HPC - KV cache ↔ storage hierarchy - gradient compression ↔ collective communication - agent workflow ↔ cluster scheduling - checkpoint ↔ GDS / distributed storage ## 4. Benchmark / Dataset / Evaluation ### Core Benchmarks for My Research ##### 1. [How Do VLMs Behave When Blind or Misled? Behavioral Evaluation of VLMs on Scientific Figures](https://arxiv.org/abs/2608.13267v1) - 阅读层级：关注 - Source: arXiv AI/ML/NLP/Vision/Robotics - 证据来源：仅摘要 - benchmark 评估什么能力：评估摘要中描述的任务能力；具体指标需打开原文确认。 - 适合用于什么研究：适合用于评测协议、指标设计或负样本构造参考；是否纳入实验需看任务贴合度。 - 可否作为实验基准：可以优先评估是否作为实验基准。 - 建议行动：use_as_eval ##### 2. [MLLM-Routed Heterogeneous Ensembles for Robust Cross-Dataset Image Classification](https://arxiv.org/abs/2608.13463v1) - 阅读层级：关注 - Source: arXiv AI/ML/NLP/Vision/Robotics - 证据来源：仅摘要 - benchmark 评估什么能力：评估 agent 规划、执行或环境交互能力。 - 适合用于什么研究：适合用于 agent evaluation / memory / long-horizon planning 相关实验。 - 可否作为实验基准：可以优先评估是否作为实验基准。 - 建议行动：use_as_eval ##### 3. [HumanTracker: Towards Comprehensive and Human-Aligned Motion Tracking Benchmark](https://arxiv.org/abs/2608.13555v1) - 阅读层级：关注 - Source: arXiv AI/ML/NLP/Vision/Robotics - 证据来源：仅摘要 - benchmark 评估什么能力：评估摘要中描述的任务能力；具体指标需打开原文确认。 - 适合用于什么研究：适合用于评测协议、指标设计或负样本构造参考；是否纳入实验需看任务贴合度。 - 可否作为实验基准：可以优先评估是否作为实验基准。 - 建议行动：use_as_eval ##### 4. [TraVEL: Trajectory-Guided Video Embedding Learning for Driving-Video Retrieval](https://arxiv.org/abs/2608.13495v1) - 阅读层级：关注 - Source: arXiv AI/ML/NLP/Vision/Robotics - 证据来源：仅摘要 - benchmark 评估什么能力：评估摘要中描述的任务能力；具体指标需打开原文确认。 - 适合用于什么研究：适合用于 agent evaluation / memory / long-horizon planning 相关实验。 - 可否作为实验基准：可以优先评估是否作为实验基准。 - 建议行动：use_as_eval ##### 5. [PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives](https://arxiv.org/abs/2608.13552v1) - 阅读层级：关注 - Source: arXiv AI/ML/NLP/Vision/Robotics - 证据来源：仅摘要 - benchmark 评估什么能力：评估 agent 规划、执行或环境交互能力。 - 适合用于什么研究：适合用于 agent evaluation / memory / long-horizon planning 相关实验。 - 可否作为实验基准：可以优先评估是否作为实验基准。 - 建议行动：use_as_eval ### Interesting Benchmarks ##### 1. [H2R-Bench: Benchmarking Human-to-Robot Manipulation Video Generation in World Models](https://arxiv.org/abs/2608.13049) - 阅读层级：关注 - Source: Hugging Face Daily Papers - 证据来源：仅摘要 - benchmark 评估什么能力：评估摘要中描述的任务能力；具体指标需打开原文确认。 - 适合用于什么研究：适合用于评测协议、指标设计或负样本构造参考；是否纳入实验需看任务贴合度。 - 可否作为实验基准：暂不作为核心基准，先保存评测协议和指标设计。 - 建议行动：save ##### 2. [Edit2TikZ: A Comprehensive and Challenging Benchmark for Scientific Figure Editing with TikZ](https://arxiv.org/abs/2608.13441v1) - 阅读层级：关注 - Source: arXiv AI/ML/NLP/Vision/Robotics - 证据来源：仅摘要 - benchmark 评估什么能力：评估摘要中描述的任务能力；具体指标需打开原文确认。 - 适合用于什么研究：适合用于多模态泛化或跨域评测设计参考。 - 可否作为实验基准：暂不作为核心基准，先保存评测协议和指标设计。 - 建议行动：skim ##### 3. [Performance Evaluation of an Adaptive Quadrature and a Double Exponential Formula Using Arbitrary-Precision Floating-Point Arithmetic](https://arxiv.org/abs/2608.13187v1) - 阅读层级：关注 - Source: arXiv Systems/HPC/GPU Data Path - 证据来源：仅摘要 - benchmark 评估什么能力：评估摘要中描述的任务能力；具体指标需打开原文确认。 - 适合用于什么研究：适合用于评测协议、指标设计或负样本构造参考；是否纳入实验需看任务贴合度。 - 可否作为实验基准：暂不作为核心基准，先保存评测协议和指标设计。 - 建议行动：skim ##### 4. [Evaluation of Clinically Steerable Retinal Image Generation from Foundation Model Latent Spaces](https://arxiv.org/abs/2608.13455v1) - 阅读层级：关注 - Source: arXiv AI/ML/NLP/Vision/Robotics - 证据来源：仅摘要 - benchmark 评估什么能力：评估摘要中描述的任务能力；具体指标需打开原文确认。 - 适合用于什么研究：适合用于评测协议、指标设计或负样本构造参考；是否纳入实验需看任务贴合度。 - 可否作为实验基准：暂不作为核心基准，先保存评测协议和指标设计。 - 建议行动：save ##### 5. [Multi-UAV Tracking Evaluation Using 5G Uplink Signals on an O-RAN ISAC Simulation Testbed](https://arxiv.org/abs/2608.10784v1) - 阅读层级：关注 - Source: arXiv Systems/HPC/GPU Data Path - 证据来源：仅摘要 - benchmark 评估什么能力：评估摘要中描述的任务能力；具体指标需打开原文确认。 - 适合用于什么研究：适合用于评测协议、指标设计或负样本构造参考；是否纳入实验需看任务贴合度。 - 可否作为实验基准：暂不作为核心基准，先保存评测协议和指标设计。 - 建议行动：skim ### Other Benchmarks - 其余 4 个只进入附录标题列表：reports/appendix/2026-08-17-benchmarks.md ## 5. GitHub / Open Source Projects ### New / Recently Active Projects ##### 1. [Paritok-official/paritok-4b-v1](https://github.com/Paritok-official/paritok-4b-v1) - Reading tier: study_code - Source: GitHub AI Research Projects (aggregator; role=code_actionability) - Published: 2026-08-16T02:39:50+00:00 - Primary track: GitHub / 开源项目推荐 - Secondary tags: 上下文压缩 / 长上下文 / 记忆, Agent / 推理 / 推理时扩展 / 规划, Compression / Reliability for AI Infrastructure, Benchmark / 数据集 / 评测, 工具库 - Grounding level: repo README - Scores: personal=0.69, global=0.62, credibility=0.88, evidence=0.69, hype_risk=0.00, feedback=0.00 - Project relevance: skyfs=0.00, schedagent=0.00, verl_infrastructure=0.00, embodied_intelligence=0.00 - What it is: Paritok-official/paritok-4b-v1: 开源项目, 方向为“GitHub / 开源项目推荐”; 主要线索: agent, agentic, compression, context window. - Problem: 它关注“GitHub / 开源项目推荐”里的 agent, agentic, compression, context window 等问题. - Method/contribution: 这是代码仓库条目; 优先检查 README, 示例, 许可证和是否有可复现实验入口. - Why important to George: Reading tier: 研读代码 editorial_priority: 0.29 按 GitHub 项目动作处理. personal: 0.69, relevance: 0.69. - Suggested action: study_code - Matched keywords: agent, agentic, compression, context window, evaluation, github, github.com, open-source ##### 2. [NousResearch/hermes-agent](https://github.com/NousResearch/hermes-agent) - Reading tier: clone_and_run - Source: GitHub AI Research Projects (aggregator; role=code_actionability) - Published: 2026-08-16T22:43:12+00:00 - Primary track: GitHub / 开源项目推荐 - Secondary tags: AI Systems / HPC / Distributed Training & Inference, Agent Runtime / RL Infrastructure / Scheduling, 工具库 - Grounding level: repo README - Scores: personal=0.81, global=0.62, credibility=0.89, evidence=0.69, hype_risk=0.00, feedback=0.00 - Project relevance: skyfs=0.00, schedagent=0.00, verl_infrastructure=0.00, embodied_intelligence=0.00 - What it is: NousResearch/hermes-agent: 开源项目, 方向为“GitHub / 开源项目推荐”; 主要线索: GPU cluster, agent, cluster, github. - Problem: 它关注“GitHub / 开源项目推荐”里的 GPU cluster, agent, cluster, github 等问题. - Method/contribution: 这是代码仓库条目; 优先检查 README, 示例, 许可证和是否有可复现实验入口. - Why important to George: Reading tier: 克隆运行 editorial_priority: 0.35 按 GitHub 项目动作处理. personal: 0.81, relevance: 0.95. - Suggested action: clone_and_run - Matched keywords: GPU cluster, agent, cluster, github, github.com, gpu, open-source ##### 3. [bytedance/deer-flow](https://github.com/bytedance/deer-flow) - Reading tier: clone_and_run - Source: GitHub AI Research Projects (aggregator; role=code_actionability) - Published: 2026-08-16T15:54:06+00:00 - Primary track: GitHub / 开源项目推荐 - Secondary tags: Agent / 推理 / 推理时扩展 / 规划, Agent Runtime / RL Infrastructure / Scheduling, 工具库 - Grounding level: repo README - Scores: personal=0.81, global=0.62, credibility=0.89, evidence=0.69, hype_risk=0.00, feedback=0.00 - Project relevance: skyfs=0.00, schedagent=0.00, verl_infrastructure=0.00, embodied_intelligence=0.00 - What it is: bytedance/deer-flow: 开源项目, 方向为“GitHub / 开源项目推荐”; 主要线索: agent, agentic, framework, github. - Problem: 它关注“GitHub / 开源项目推荐”里的 agent, agentic, framework, github 等问题. - Method/contribution: 这是代码仓库条目; 优先检查 README, 示例, 许可证和是否有可复现实验入口. - Why important to George: Reading tier: 克隆运行 editorial_priority: 0.35 按 GitHub 项目动作处理. personal: 0.81, relevance: 0.94. - Suggested action: clone_and_run - Matched keywords: agent, agentic, framework, github, github.com, long-horizon, multi-agent, open-source ### Paper-linked Repos ##### 1. [deepseek-ai/DeepSeek-OCR](https://github.com/deepseek-ai/DeepSeek-OCR) - Reading tier: study_code - Source: GitHub AI Research Projects (aggregator; role=code_actionability) - Published: 2026-01-27T03:45:14+00:00 - Primary track: GitHub / 开源项目推荐 - Secondary tags: Agent / 推理 / 推理时扩展 / 规划, AI Systems / HPC / Distributed Training & Inference, Benchmark / 数据集 / 评测, Compression / Reliability for AI Infrastructure, 工具库 - Grounding level: repo README - Scores: personal=0.65, global=0.45, credibility=0.89, evidence=0.69, hype_risk=0.00, feedback=0.00 - Project relevance: skyfs=0.00, schedagent=0.00, verl_infrastructure=0.00, embodied_intelligence=0.00 - What it is: deepseek-ai/DeepSeek-OCR: 开源项目, 方向为“GitHub / 开源项目推荐”; 主要线索: compression, environment, eval, github. - Problem: 它关注“GitHub / 开源项目推荐”里的 compression, environment, eval, github 等问题. - Method/contribution: 这是代码仓库条目; 优先检查 README, 示例, 许可证和是否有可复现实验入口. - Why important to George: Reading tier: 研读代码 editorial_priority: 0.11 按 GitHub 项目动作处理. personal: 0.65, relevance: 0.69. - Suggested action: study_code - Matched keywords: compression, environment, eval, github, github.com, image, inference, open-source ##### 2. [rednote-machine-learning/RedKnot](https://github.com/rednote-machine-learning/RedKnot) - Reading tier: study_code - Source: GitHub AI Research Projects (aggregator; role=code_actionability) - Published: 2026-07-10T06:18:48+00:00 - Primary track: GitHub / 开源项目推荐 - Secondary tags: AI Systems / HPC / Distributed Training & Inference, 上下文压缩 / 长上下文 / 记忆, 其他亮点, GPU-Centric I/O / Networking / Storage, 工具库 - Grounding level: repo README - Scores: personal=0.61, global=0.36, credibility=0.88, evidence=0.69, hype_risk=0.00, feedback=0.00 - Project relevance: skyfs=0.00, schedagent=0.00, verl_infrastructure=0.00, embodied_intelligence=0.00 - What it is: rednote-machine-learning/RedKnot: 开源项目, 方向为“GitHub / 开源项目推荐”; 主要线索: alignment, attention, github, github.com. - Problem: 它关注“GitHub / 开源项目推荐”里的 alignment, attention, github, github.com 等问题. - Method/contribution: 这是代码仓库条目; 优先检查 README, 示例, 许可证和是否有可复现实验入口. - Why important to George: Reading tier: 研读代码 editorial_priority: 0.10 按 GitHub 项目动作处理. personal: 0.61, relevance: 0.68. - Suggested action: study_code - Matched keywords: alignment, attention, github, github.com, inference, long-context, open-source, serving ##### 3. [microsoft/MInference](https://github.com/microsoft/MInference) - Reading tier: clone_and_run - Source: GitHub AI Research Projects (aggregator; role=code_actionability) - Published: 2026-04-08T08:04:38+00:00 - Primary track: GitHub / 开源项目推荐 - Secondary tags: 上下文压缩 / 长上下文 / 记忆, 模型架构, AI Systems / HPC / Distributed Training & Inference, 其他亮点, 工具库 - Grounding level: repo README - Scores: personal=0.64, global=0.48, credibility=0.88, evidence=0.69, hype_risk=0.00, feedback=0.00 - Project relevance: skyfs=0.00, schedagent=0.00, verl_infrastructure=0.00, embodied_intelligence=0.00 - What it is: microsoft/MInference: 开源项目, 方向为“GitHub / 开源项目推荐”; 主要线索: attention, github, github.com, inference. - Problem: 它关注“GitHub / 开源项目推荐”里的 attention, github, github.com, inference 等问题. - Method/contribution: 这是代码仓库条目; 优先检查 README, 示例, 许可证和是否有可复现实验入口. - Why important to George: Reading tier: 克隆运行 editorial_priority: 0.13 按 GitHub 项目动作处理. personal: 0.64, relevance: 0.65. - Suggested action: clone_and_run - Matched keywords: attention, github, github.com, inference, long-context, open-source, release, sparse attention ### Evergreen Toolkits - 今日无需要重复推荐的常青工具库。 ## 6. Scholar Radar - Jeff Dean: focus=ai_systems_hpc, distributed_systems, machine_learning_systems; last_verified=2026-07-18 - Richard Sutton: focus=rl, agent_rl_infrastructure; last_verified=2026-07-18 - Torsten Hoefler: focus=ai_systems_hpc, gpu_data_path_storage, compression_reliability; last_verified=2026-07-18 - Pieter Abbeel: focus=embodied_world_models, rl; last_verified=2026-07-18 - Shunyu Yao: focus=agent_rl_infrastructure, agents; last_verified=2026-07-18 - 孙凝晖: focus=ai_systems_hpc, hpc; last_verified=2026-07-18 - 赵海睿: focus=agent_rl_infrastructure, ai_systems_hpc; last_verified=2026-07-18 ## 7. University / Lab Radar - [Identifying Interactions at Scale for LLMs](http://bair.berkeley.edu/blog/2026/03/13/spex/) - 学校 / 实验室：UC Berkeley - 类型：project - 为什么值得关注：institution_signal 0.96，authority_score 0.96 - 与我的研究方向关系：AI 系统 / HPC / 分布式训练与推理，personal 0.87 - 建议行动：watch - [Spatial Memory Agent: Experience-Grounded Procedure Memory for Spatial Intelligence](https://arxiv.org/abs/2608.12743) - 学校 / 实验室：Hugging Face - 类型：paper - 为什么值得关注：institution_signal 0.96，authority_score 0.96 - 与我的研究方向关系：Agent / 推理 / 推理时扩展 / 规划，personal 0.86 - 建议行动：watch - [Adaptive Parallel Reasoning: The Next Paradigm in Efficient Inference Scaling](http://bair.berkeley.edu/blog/2026/05/08/adaptive-parallel-reasoning/) - 学校 / 实验室：UC Berkeley - 类型：dataset - 为什么值得关注：institution_signal 0.96，authority_score 0.96 - 与我的研究方向关系：Agent / 推理 / 推理时扩展 / 规划，personal 0.85 - 建议行动：skim - [AVA-Encoder: Towards Agent-Native Video Representation Learning](https://arxiv.org/abs/2608.12313) - 学校 / 实验室：Hugging Face - 类型：paper - 为什么值得关注：institution_signal 0.96，authority_score 0.96 - 与我的研究方向关系：Agent / 推理 / 推理时扩展 / 规划，personal 0.82 - 建议行动：watch - [H2R-Bench: Benchmarking Human-to-Robot Manipulation Video Generation in World Models](https://arxiv.org/abs/2608.13049) - 学校 / 实验室：Hugging Face - 类型：paper - 为什么值得关注：institution_signal 0.96，authority_score 0.96 - 与我的研究方向关系：Benchmark / 数据集 / 评测，personal 0.82 - 建议行动：watch ## 8. Company Research Radar - Stanford University: focus=ai, systems, robotics; last_verified=unverified - MIT: focus=ai_systems_hpc, robotics; last_verified=unverified - UC Berkeley: focus=systems, ai, robotics; last_verified=unverified - Carnegie Mellon University: focus=systems, robotics, ai; last_verified=unverified - Tsinghua University: focus=ai_systems_hpc, ai; last_verified=unverified - Institute of Computing Technology, CAS: focus=ai_systems_hpc, distributed_systems; last_verified=unverified - NVIDIA Research: focus=gpu_data_path_storage, ai_systems_hpc, embodied_world_models; last_verified=unverified - Google DeepMind: focus=ai, embodied_world_models, rl; last_verified=unverified ## 9. Associations / Awards / Fellows / Leadership ### Associations - ACM: focus=computer_science, ai_systems_hpc; last_verified=2026-07-18 - IEEE: focus=computer_science, electrical_engineering; last_verified=2026-07-18 - IEEE Computer Society: focus=ai_systems_hpc, gpu_data_path_storage; last_verified=2026-07-18 - AAAI: focus=ai, agent_rl_infrastructure; last_verified=2026-07-18 - CCF: focus=computer_science; last_verified=2026-07-18 - USENIX: focus=systems, security, storage; last_verified=2026-07-18 - SIAM: focus=hpc, scientific_computing; last_verified=2026-07-18 - ACL: focus=nlp; last_verified=2026-07-18 ### Awards & Notable Papers - 今日无高相关顶会精选。 ## 10. Notable Conference and Journal Papers - NeurIPS: focus=not specified; last_verified=2026-07-18 - ICML: focus=not specified; last_verified=2026-07-18 - ICLR: focus=not specified; last_verified=2026-07-18 - SOSP: focus=not specified; last_verified=2026-07-18 - OSDI: focus=not specified; last_verified=2026-07-18 - FAST: focus=not specified; last_verified=2026-07-18 - SC: focus=not specified; last_verified=2026-07-18 - SIGCOMM: focus=not specified; last_verified=2026-07-18 ## 11. Evergreen Classics ### 1. [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401)（2020） - 作者：Patrick Lewis、Ethan Perez、Aleksandra Piktus、Fabio Petroni、Vladimir Karpukhin、Naman Goyal、Heinrich Kuttler、Mike Lewis 等 - topic_tags：context_compression、long_context、nlp - 关联方向：Context Compression / Long Context / Memory、NLP - 为什么经典：RAG 把参数记忆和外部检索连接起来，是理解今天检索增强、agent memory、记忆有效性和上下文预算取舍的关键参照。 - 今日新论文继承了什么问题：今天的相关条目 延续了经典工作里的核心问题：有限上下文、外部记忆与状态复用如何支撑更长程的推理。 - 它挑战了什么经典假设：它挑战的是静态检索、固定窗口或只读记忆的假设，转向会随新证据更新的工作记忆和缓存管理。 - 它推进到什么新场景：新场景从语言建模推进到 agent memory、动态 workflow 和长上下文服务系统。 - 预备知识：了解 seq2seq、dense retrieval 和生成式问答。 ## 12. Feedback-Aware Recommendations - No explicit feedback signal yet; using cold-start research profile. ## 13. Source Health - OpenReview：错误（0 条） - 返回内容为空或不是合法 JSON: line 1 column 1 (char 0) - GitHub AI Research Projects：time budget exhausted（23 条） - 时间预算已耗尽 after 23 items - MIT CSAIL News：超时（0 条） - timeout after 25s - The Batch by DeepLearning.AI：错误（0 条） - 403 Client Error: Forbidden for url: https://www.deeplearning.ai/the-batch ## 14. Collection Notes - Generated at: 2026-08-16T22:51:31.297616+00:00 - Source count: 31 - Raw item count: 692 - Dedup item count: 563 - API requests total: 7 - API requests by provider: deepseek:6, kimi:1 - Cache hits: 1 - Cache misses: 5 - Benchmark appendix: reports/appendix/2026-08-17-benchmarks.md - Report path: reports/daily/2026/08/2026-08-17.md - 上一份报告链接：reports/daily/2026/08/2026-08-16.md