Agent-Computer Observation Interfaces Enable Dynamic Computer Use (AOI / DynaCU-Bench)
动态计算机使用：观察接口+流式+动作接地
Multi-agent Framework for Time-Sensitive Complementary Collaboration in Minecraft (TickingCollabBench)
Minecraft时间敏感协作基准：同步vs异步、wall-clock
SentinelBench: A Benchmark for Long-Running Monitoring Agents
长时监控agent基准：speed_factor+wall-clock
Benchmarking Living-Screen-Native GUI Agents on Short-Video Platforms (LivingScreen)
短视频平台原生动态屏幕GUI agent基准
DuplexSLA: A Full-Duplex Spoken Language Model with Synchronized Speech, Language, and Action
全双工语音-语言-动作同步模型，共享时钟+动作通道
Omni-DuplexEval: Evaluating Real-time Duplex Omni-modal Interaction
全双工全模态实时交互评测
GameWorld: Towards Standardized and Verifiable Evaluation of Multimodal Game Agents
游戏agent状态可验证评测基准，ECCV 2026
τ-Voice: Benchmarking Full-Duplex Voice Agents on Real-World Domains
全双工语音agent真实场景基准：tick编排器+音频-时间解耦
Video Streaming Thinking: VideoLLMs Can Watch and Think Simultaneously (VST)
边看边想：流式CoT+推理摊销，ECCV 2026
Gaia2: Benchmarking LLM Agents on Dynamic and Asynchronous Environments (+ ARE)
动态异步环境agent基准，ICLR 2026 Oral
Training-Time Action Conditioning for Efficient Real-Time Chunking
训练期动作条件化：模拟延迟以支持高效实时分块（Physical Intelligence）
Real-Time Reasoning Agents in Evolving Environments (RealtimeGym / AgileThinker)
实时推理agent：按token计时的推理成本
StreamingThinker: Large Language Models Can Think While Reading
边读边想：流式CoT+并行KV缓存
SHANKS: Simultaneous Hearing and Thinking for Spoken Language Models
边听边想：增量推理+提前工具调用
STITCH: Simultaneous Thinking and Talking with Chunked Reasoning for Spoken Language Models
边说边想：分块推理+播放窗口
Real-Time Execution of Action Chunking Flow Policies (RTC)
实时动作分块：推理时插值实现延迟补偿（Physical Intelligence）
Win Fast or Lose Slow: Balancing Speed and Accuracy in Latency-Sensitive Decisions of LLMs
速度-精度权衡：延迟敏感决策的倒U型关系
VideoGameBench: Can Vision-Language Models complete popular video games?
VLM电子游戏实时性基准
Optimizing Anytime Reasoning via Budget Relative Policy Optimization (AnytimeReasoner)
Anytime推理优化：BRPO预算采样+可验证稠密奖励
OneTwoVLA: A Unified Vision-Language-Action Model with Adaptive Reasoning
自适应推理VLA：决策token+事件触发的推理摊销
HiRT: Enhancing Robotic Control with Hierarchical Robot Transformers
分层机器人Transformer：slow-fast异步VLM架构+潜变量缓冲
Robotic Control via Embodied Chain-of-Thought Reasoning (ECoT)
具身CoT推理：分块推理导致的阻塞式推理延迟问题
FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference
单卡 30Hz 控制频率、平滑异步执行：FlashVLA 用「流式动作缓冲区」解决 VLA 模型部署时的延迟和异步执行不稳定这两大痛点。
MemToC: Benchmarking Memory-Tool Conflict Resolution in Large Language Models
工具返回的结果和模型自己记的不一样时该信谁？MemToC 用 6504 个受控场景量化这个「记忆-工具冲突」问题，结论是：模型几乎总是无脑信工具。
GraphMemix: Query-Aware Evidence Forests for Long-Term Multimodal Agent Memory
多模态 agent 的长期记忆不该是「离线摘要」或「粗暴相似度检索」，而应该按查询构建一片「证据森林」。
Understanding Evolution Strategies for LLM Reasoning: Broader Reasoning Coverage than GRPO
进化策略（ES）不是 GRPO 的「省内存平替」，而是一条能拓宽推理覆盖面的独立后训练路线——理论+实验双重证据。
Performance Foundations of Parallel & Distributed RLM Training
RLM 训练的并行策略分类法与系统性工程指南
Adaptive Reasoning Allocation in Agentic AI
过度推理和推理不足是同一个「推理误分配」问题的两面
Persona-Execution Separation for Governed Agents
人设可漂移、执行可审计的治理型 agent 架构模式
Instruct-to-Act: Decoupling Planning and Control
VLM 高层慢规划 + 世界模型控制器高频快执行
Contract-Centered Architecture for Agentic Runtimes
企业级 agent 部署的组织契约架构提案（无实验结果）
UrbanGround: Spatial Agency in Real-Scale City
1:1还原香港城市沙盒里测试空间智能体的组合失效
PAWBench: Probabilistically Aligned World Modeling
视频生成世界模型的概率对齐评测基准
GameWAM: World-Action Model for Video Games
同时生成未来画面与可执行键鼠动作序列的游戏智能体
PILOT: Live Self-Improvement for Long-Horizon Agents
监督者-执行者架构实现运行中实时纠偏与技能沉淀
Self-OPD: Teacher-free On-Policy Distillation
flow matching 模型自我探索、自我打分、自我蒸馏
TTPO: Test-Time Policy Optimization
无标签测试时训练，用投票分歧做不对称蒸馏+惩罚
ACE Lens: Unifying Agentic Data Generation
用 Accuracy-Complexity-divErsity 三轴透镜统一 agentic 数据生成范式
One Policy, Many Embodiments: Unified Camera-Centric Action Geometry Pre-training for Heterogeneous Embodied Manipulation
相机中心统一动作几何预训练实现跨本体泛化
StreamPI: Streaming Multimodal Temporal Modeling for Vision-Language-Action Models
零新增参数为单帧VLA装上流式时序推理能力
$R^3$: Training Robots to Reason in Natural Language via Reinforcement Learning
用自然语言推理训练机器人操作策略的测试时计算机制
One Symptom, Three Levers: A Critical Review of On-Policy Self-Distillation
On-Policy Self-Distillation坍缩问题的三杠杆综述
Spectral Allocation: Why Muon Outperforms Adam, and How to Improve Muon
频谱探测解释Muon优越性并提出SAMuon优化器
AsymSpec: Context-Asymmetric Speculative Decoding for Agentic LLMs
非对称上下文投机解码提升agentic LLM推理效率
Reflection Steering: Disentangling Reflection from Reasoning in Activation Space for Token-Efficient Inference
分离反思方向以实现token高效推理
Reconstructing the Right Episode: Evaluating Interleaved Conversational Memory Beyond Long Context
扁平交错对话线程中的情节完整性记忆基准SCALE-QA
AWM: Answerable Working Memory for Long-Document VQA Agents
长文档VQA agent工作记忆的可回答性评测与GRPO优化
A Storage-Retrieval Gap in Parametric Knowledge Graph Memory
参数化知识图谱记忆的存储-检索鸿沟
Learning What to Share and What to Personalize: Hierarchical Strategy Co-Evolution for Agent Memory
记忆管理策略的全局共享与个性化分层共演化
Repair or Resample? Rethinking Failure Debugging in LLM Multi-Agent Systems
多智能体系统修复方法到底是因果修复还是随机侥幸
TOPAS: Workflow-Aware Prefix-State Scheduling for Multi-Agent LLM Serving
联合决定缓存哪些前缀与调度哪些请求，降低多智能体服务延迟
JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution
把agent harness本身当作可训练、可即时生成的产物
SwarmWorld: Stigmergic technological evolution in societies of language-model agents
同质化LLM agent能否靠物理痕迹自发分化社会分工
Tunable Tool-Call Rates in LLM Agents via Representation Steering
用残差流中的单一线性方向调节工具调用率，无需训练
AgentSpec: Speculative Decoding for Batch Inference of LLM Agents
针对agent工作流结构重新设计投机解码，解决大batch场景下的失灵问题
Parason: Revealing Subtask and Trial Parallelism in LLM Reasoning
拆解长推理链中的两种并行结构，PA-GRPO训练实现1.7x推理加速
Latent Action as Intention Enables Efficient Future Imagination for World Action Models
LAWA：以潜在动作作为意图表征，跳过未来视频生成实现高效"未来想象"
Benchmarking LLM Judges for Voice-Agent Evaluation: Do Text-Based Judges Miss What Matters?
文本裁判在自然度和轮次交替质量上明显落后于音频原生裁判和人类
FireRedAudio: An Open Foundation Model for Unified Audio Understanding and Generation
统一离散音频token，单一自回归架构同时做音频理解与生成
Learning to Act While Waiting: Latency-Aware Policies for Real-Time Embodied Agents
延迟感知策略：双策略头+门控函数让具身agent在实时约束下优雅降级
The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents
模型间升级/降级切换的成本-质量代价，交接税量化
Game2World Engine: Unlocking In-the-Wild Gameplay Videos for World Model Training
自动提取/移除游戏UI叠加层，清洁数据提升世界模型效果
Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses
经验-工作记忆递归自改进架构，在37个模型-基准组合中35个取得提升
MemUse: Moving Memory Evaluation from Direct QA to Natural Integration in Long-Term Human-AI Conversation
4个月真实部署证明Direct QA分数与用户满意度脱节，自然整合能力才是关键
From Seeing to Acting: Smart Glasses as First-Person Intelligence Platforms
首个智能眼镜统一综述框架，L0-L5能力分级
AgentRoom: Concurrent Multi-Agent Coding in a CRDT-Backed Shared Workspace
CRDT共享工作区让编码agent真正并发协作，证明起效的是协调机制而非并行度
Adaptive Influence Graphs for Failure Attribution in Multi-Agent Systems
用结构化影响图取代原始日志，两阶段agentic框架在Who&When基准刷新SOTA
WeMM-Embedding: WeChat Multi-Modal Embedding Technical Report
微信通用多模态嵌入模型，MMEB-v2新SOTA已规模化部署
On-policy Distillation with Verifiable Reward
无新超参数融合RLVR与OPD，ReLU门控对齐蒸馏信号与任务成功
When "Must" Becomes "Maybe": Constraint Weakening in LLM Agent Workflows
交接压缩把安全阻断器悄悄降格为仅供参考的信息
Automata from Agent Traces: Failure and Next-Step Prediction
把agent trace坍缩成有限状态机，支撑失败预测与下一步预测
CAFE: Self-Improving Search Agents Need Co-Evolving Feedback
agent与critic交替共同演化，只改进单方会停滞
Knowing When to Ask for Help: Bayesian Self-Escalation in Hierarchical LLM Agents
生成中主动升级的贝叶斯最优停止理论，1/√n regret保证
Markets, Not Planners: Decentralized Orchestration of LLM Agents with Private Information
用劳动力市场机制取代中心化planner编排LLM agent
The Empire, Long Divided, Must Unite: Architectural Convergence in Three LLM Agent Harnesses
三个理念相反的coding agent harness收敛到同一架构
Meta$^n$: Recursive Self-Improvement through Emergent Depth
元层级固定、递归作用于自身输出的自改进架构
From State to Action: OODA-Tool for Reliable Multi-Turn Tool Use
OODA-Tool 用 Observe-Orient-Decide-Act 闭环分离状态维护与动作生成，缓解多轮工具调用中的状态-动作竞争
When May an Agent Stop? Evidence-Carrying Termination for Tool-Using LLMs
ECT 要求终止声明绑定可确定性回放的证据，在受控故障测试中把不安全过早终止降为0
Logic-VLA: A Temporal Logic Conditioned Vision-Language-Action Model
用STL时序逻辑规约条件化VLA策略，推理时按需切换安全约束而无需为每条规约单独训练
RAG Deserves an Index: Why Ingest-Time Compilation Beats Query-Time Interpretation
RAG如同全表扫描：把语义理解从查询时挪到摄入时编译成索引，检索准确率85.2%仅用2.2k token优于分块法72.5%@16.3k token
XPerf: Benchmarking LLM Serving Systems for Agentic AI Workloads
用细粒度trace replay解决agentic负载非确定性问题，实现LLM serving系统间的可复现压测与瓶颈定位
Is Multimodal Speculative Decoding Ready for Diffusion-Based Parallel Drafting?
综述+实证诊断：扩散式并行起草在多模态（VL/Video-Language/Audio/VLA）投机解码中仍是空白
LiLiCorr: Lightweight Likelihood Correlation of Parallel Drafts for Speculative Decoding
用轻量 in/out 向量余弦相似度打分，修复扩散式区块草稿头的联合不连贯问题，接受长度提升9-19%
When Failures Propagate: Causal Failure Attribution in Agentic RAG
干预式基准揭示：agentic RAG 的事后失败归因在 hop 1 准确率0.91，但一旦故障传播到 hop 2/3 就完全失灵（0.00）
CAS: Conformalized Agentic Search via Adaptive Retrieval and Policy Weighting
用 Conformal Prediction 同时约束检索截断与 GRPO 训练置信度，缓解 RL 搜索 agent 的证据丢失与过度自信问题
Tree-of-Concerns: Hierarchical Multi-Agent Debate for Unstated-Limitation Extraction in Scientific Critique
用专属视角的怀疑者人格辩论树+合议庭复核，从科学论文中挖掘未言明的局限性，精确率提升79%
GRAFT: Adaptive DLM-Based Draft Tree Construction with Target-Distilled Edge Scoring
用目标模型蒸馏的父子边打分+状态感知动态预算，让扩散语言模型drafter的树形投机解码提速2.13x-6.36x，开销<0.5ms
Applying Anthropic Primitives at Large Enterprises: Harness Paradigm for Knowledge Work
把 coding-agent harness 当作企业基础设施而非编码工具，用统一 harness + 凭据范围化工具 + 外置授权逻辑，让"审阅新构建的东西"收敛为"读一份指令文件"。
PrimeAgentOrchestrator: Memory-Primed Agent Spawning for Personal AI Infrastructure
四个月生产部署复盘：如何在 spawn 时给 Claude Code 实例注入跨后端记忆，绕开每次会话的上下文归零问题
ForeTime-VLA: Causal Future-Token Distillation from a World Action Model for Conveyor-Belt Manipulation
将世界动作模型的未来预测蒸馏成因果 token，让 π0.5 策略在不牺牲实时性的前提下预判移动物体的接触时机
AgentMercury: 从商业场景规模化合成可验证 Agent 环境
先造一个持久化商业世界，任务从中自然涌现——RL训练后域内域外双双大涨，环境构造能力本身也能被学会（3.3%→83.3%）
Automated Trajectory Evaluation for Mobile Agents via Step-Level Consequence Reasoning and Aggregation
CRATE用两阶段VLM-as-judge做移动智能体轨迹评测：步骤级后果推理消解上下文过载，AndroidWorld F1达0.833，并衍生出安全评测版CRATE-S
Temporal Validity on Real Software Histories: Eliminating Stale-Fact Errors in Code-Assistant Memory over GitHub Fixes
真实 GitHub 修复数据验证：确定性三元组覆盖式记忆 MemStrata 把 RAG 的陈旧事实错误率从 36-38% 压到接近 0
DreamBench-SWE: A Multi-Session Memory-Hygiene Benchmark for Software Agents
严谨预注册审计揭示：软件agent跨会话记忆最好方案通过率仅53.89%，无记忆基线仅11.67%
2027 上半年智能体方向判断（全文）
AgentTeam-Shared-Knowledge 内部报告全文转载，含目录锚点，适配手机阅读
Three-Dimensional Typology of Agency in LLM Systems
Agency三维类型学：从能力、自主性、意图性三轴统一定义LLM agent，厘清术语混乱
PolicyGuide: Workflow-Level Policy Compliance for AI Agents
PolicyGuide：工作流级政策合规，agent执行前自动校验操作是否符合组织政策约束
MidTool: Mid-Training for Tool Use Capability
MidTool：预训练中期插入工具使用训练，比后训练更高效地习得工具调用能力
Cross-Task Skill Transfer for Tool-Using Agents
跨任务技能迁移：从已有工具调用经验中抽取可复用技能，迁移到新任务降低样本需求
SAPO: Single-Rollout Agentic Policy Optimization
SAPO：单轨迹agentic RL，无需多轨迹对比即可稳定训练，降低采样成本
Learning When to Think: Adaptive Reasoning Modes in LLMs
自适应推理模式：LLM何时快想何时慢想，动态切换提升效率而不损准确率
Optimal Skill Selection: Bicriteria Guarantees for Agent Skill Composition
最优技能选择：双目标保证——覆盖率与效率的可证明权衡，agent技能组合的理论基础
Thinkingbox: Stateful Business Workflow Benchmark for LLM Agents
Thinkingbox：有状态商业工作流benchmark，测LLM能否在复杂业务流程中维护状态并完成多步任务
ReCache: KV Cache Reuse for Tool-Augmented Agents
ReCache：工具增强型agent的KV缓存复用，减少重复工具调用的推理开销
Outcome Monitors: Silent Tool Failure Detection in Agents
结果监测器：沉默式工具失败检测，agent无需感知工具内部即可识别失败
MemTrapBench: Cognitive Traps in LLM Memory Use
认知陷阱基准：测量LLM记忆使用中的虚假记忆、过度依赖、混淆等认知偏差
MCB: Memory Commitment Boundary for Hybrid Memory Agents
记忆承诺边界：LLM何时应从检索转向参数记忆，混合记忆决策框架
Beyond Memory Majority: Latent-Source Reasoning for Knowledge Conflict
超越记忆多数投票：隐源推理解决LLM知识冲突，判断该信任哪个来源
StateMemBench: Benchmarking Memory State Tracking in LLM Agents
记忆状态追踪基准：动态环境下agent能否准确维护多实体状态
Represented but Ignored: Causal Analysis of Prosodic Underuse in LLMs
韵律信息被表征但被忽略：因果分析揭示LLM系统性低估语调信号
Hear2Act: Prosody-Driven Task Decision Benchmark
语音韵律→任务决策基准：8种韵律变体×多任务，测LLM能否从语调听出用户意图
The Embodiment Gap in Robot Foundation Models
具身鸿沟调查：模型泛化≠能在特定机器跑，两轴图显式化可复用部分与新本体适配工作
Reinforced Planning with Latent World Models
搜索规则也可强化学习：RP1 全离线学改进计划，roll-out 少 1000×，快 67×
SkillGate: Training In-Policy Skill Selection in Long-Horizon Agents
SkillGate 双信用通道修 selector credit starvation：9B 策略 40.8%→53.2%
RTPO: Reverse-Turn Policy Optimization for Stabilizing Agentic RL Training
反向树+逆序更新：从结构上拆掉多轮 agentic RL 的上下文错配/信用弱/异步漂移
LT-Mem: Volatility-Aware Spatio-Temporal Memory for Lifelong Scene Understanding
波动感知时空记忆演化：确定性证据评分保跨会话身份，Tri-Memory 存状态+历史
MemFuse: Multi-Source Memory Fusion from Fragmented Observations
多源碎片观察融合成连贯情景记忆：事件层原子+簇层融合，因果图保溯源
Towards Reversible Forgetting: Managing Obsolete Knowledge in Continual Enterprise AI Agents
可逆遗忘：企业 agent 记忆三态（活跃/休眠/退役），条件回归时可复活，滞回控制防振荡
Beyond the Transcript: Detecting Covert Coordination in Latent Multi-Agent Communication
VLA 监测多 agent 隐状态共谋通道：同质 agent 检测 AUROC 0.993
Bayesian Partner Modelling enables Adaptive Replanning for LLM Coordination
贝叶斯跟踪伙伴技能，只在矛盾时打断重规划——信念-行动差距大幅收窄
FM-Bench: A Benchmark for Long-Horizon Management with Competing Agents
LLM 当 20 年足球俱乐部经理：26 工具 340-400 决策点，15 个前沿模型首个头对头长程评估
Test-Time Scaling in the Wild: Why Exploitation, Not Exploration, Is the Bottleneck
开放任务 TTS 瓶颈在利用不在探索：奖励模型与真实质量相关仅 ρ≈0.12
Beyond LLM-Based Reasoning: Lightweight GNNs for Agent Failure Attribution
AFANet 轻量 GNN 做多 agent 失败归因：近零推理成本匹配或超越 LLM 基线
One Gate Is Not Enough: Composing Stateful Pre-Action Controls for Agentic AI
多闸门预执行控制会互相干扰：补救算子不交换，补救顺序是控制面语义而非实现细节
Looped Language Models Improve Compositional Tool Calling
循环 LM 做组合式工具调用：循环深度越大多步准确率越高，自适应推理算力更优
Alignment Is All You Need: Instruction-Free Training for General Audio-Language Models
音频-语言模型只用对齐阶段：双冻结只训投影器，匹配或超越重 post-training 基线
X2Streaming-TTS: Causal Token-Level Text-to-Speech from Streaming Text
真·token 级流式 TTS：因果承诺+语音状态继承，低延迟对话系统不再等句末文本
GigaBrain-0.7: Scaling Embodied Foundation Models to Emergent Capabilities with a Three-System Architecture
GigaBrain-0.7：三系统架构统一理解/预测/动作 + 3.7 万小时异构数据 + 单阶段对齐，跨本体泛化
The Null Token Knows: Reducing Message-Free Hallucination in ASR and NMT
Null Token Knows：ASR/NMT 的 message-free 幻觉抑制，null token 分数携带可用弃权信号
SEER: Long-Context Reasoning via Selective Visual-Text Compression
SEER：视觉扫描选相关图、只在需要处回文本检索，视觉压缩效率+文本推理精度结合
TRCA: Transition-wise Rubric Credit Assignment for Long-horizon LLM Agents
TRCA：从 action-induced transition 直接导出 step 级监督，Evidence/Execution/Invalidity 三 rubric，无需成功锚
MELD: A Protocol for Merging Knowledge Across Distributed Agentic Memories
MELD：联邦 agent 记忆一致性协议，五结果接纳程序 + 三信号裁决 + 可审计 Patch + CRDT
HyperSkill: Self-Evolving LLM Agents via Hypergraph-Structured Skill Memory
HyperSkill：超图结构技能记忆，subtask+技能双节点，双路径检索按共现排序，结构知情维护
QUMem: Personalized Memory for Query-Conditioned User-State Inference in LLM Agents
QUMem：语义连续性切可变长 episode、同交互多信息拆独立单元，查询条件的用户状态推断
BRA-Audit: Budgeted Runtime Auditing for LLM Multi-Agent Systems via Cumulative-Exposure Audit-Point Placement
BRA-Audit：把多 agent 审计调度建模为固定预算下的审计点放置，按累积暴露度贪心优化
Emergent Misaligned Communication in Long-Horizon Multi-Agent LLM Commerce
Vending-Bench Arena：13 个 LLM 一年期经济模拟，2583 封 agent 间邮件 12.6% 含错位 speech-act
The Hallucination Snowball: Modeling Error Propagation as State Transitions in Multi-Agent LLM Pipelines
幻觉雪球：顺序多 agent 管道交接不验证，幻觉逐跳变形逃逸，四级逃逸概率 24.6%/48.3%/89.3%
CAPO: Constraint-Aware Prompt Optimization for LLM Agents
CAPO 原-对偶方法把系统提示词当可优化变量，在操作约束下优化，无需领域监督数据
When Agentic Executions Fail: Detecting and Localizing Runtime Faults from Telemetry
AGENTCHAOSBENCH：工具/模型/guardrail/agent 边界注入 10 类故障，从遥测检测定位运行时失败
AstronOS: A Unified Execution Model and Runtime for Long-Horizon Agentic Systems
work item 跨调用持久身份+版本化权威状态，结果经校验才推进——长程 agent 的统一执行模型与跨会话交接
From Sequence to Structure: Relational Uncertainty Propagation for LLM Agents
RUPA：把执行历史建成轨迹图，不确定性沿时序/语义依赖传导，捕获根因在多步之前的失败
ParaTempo: Efficient Parallel Reasoning via Temporal Confidence
异步并行推理：temporal confidence 单一信号驱动剪枝/退休/续跑，branch-level 延迟决策与 duplex 推理层同构
DuplexGen: Decoupling Content, Timing, and Acoustics for Synthetic Dialogue Speech
今日 duplex 头条——双 full-duplex 模型实时互听执行剧本，对话 timing 交互驱动地自然涌现，内容/时序/声学三层解耦
Reflex: Enabling Fast and Predictive Vision-Language-Action Models for Reaction-Critical Manipulation
反应关键操作评测补位：ReflexBench 六动态任务+可配延迟；ReflexVLA 潜在未来预测+多帧时序融合增强时序推理，批式视觉编码降延迟
MobileMem: Learning from a Year of Mobile Experiences
设备端长时记忆 benchmark：一年级移动经验为基底，knowledge-grounded 合成时序一致长程轨迹，文本+多模态覆盖多跳/时序推理与隐式偏好
SimpleOPD: Simple Tokenizer-Agnostic On-Policy Distillation for Long-Context Reasoning
tokenizer-agnostic on-policy 蒸馏：共享文本空间只对齐同跨度 token，reference KL+终止 token 掩码稳住分布；Intern-S2-Preview ProofBench +21.2 超 Gemini-2.5-Pro
GRPO Beyond English: A Large-Scale Study of GRPO in Non-English and Multilingual Settings
大规模多语言 GRPO 实证：母语推理训练离英语差距小、跨语言迁移强，但特定模型-语言组合会严重回归，需配广泛评测
Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning
知识（FFN）与推理（Self-Attn）解耦：Mobius 全局共享 Memory 反复查询；7B 用 62.6% 数据达同分，Intern-S2-Mobius 同分下近 4x 端到端加速
Amplified Does Not Mean Predictive: Reasoning Behaviors in Thinking Models
Behavioral Lift 揭示 Amplification-Lift Gap：thinking 模型强放大自我修正/不确定承认，但最高 lift 是置信校准——训练放大表面形式而非最强正确信号
Think in Latent, Explain in Language: Self-Explainable Latent Reasoning
SELR 单模型双目标：Answer Loss 优化潜在轨迹出答案，CoT Loss 训练同模型把隐表示解码回可读步骤，隐式推理不再黑箱
VoiceChat-TTS: A Low-Latency Continuous Speech Synthesis Model for Interactive Agents
语音侧最贴近 duplex 的进展：LLM 文本 token 流直接驱动、控制 token 显式中断、无输入输出静音、mid-utterance 打断不重置 KV cache
Split the Labor: Separating Evidence Interpretation from Decision Aggregation
把「从多源下结论」拆成证据解释与决策聚合两半；四字段证据元组固定两半，pooled 校准 log-likelihood 比率修复 count-scale drift
Wrong but Useful: Trajectory Value Beyond Answer Correctness in Multi-Agent Messages
错误答案常含可用结构：DHD replay 协议测轨迹价值，5 benchmark×2 模型族全现 wrong-helpful，改变最终正确性的错误消息中 >40% 有益
Clearing the Fog: Towards Installing and Refining Proactive Exploration Capabilities in LLM Agents
给 agent 灌输主动探索能力：探索密集轨迹合成缓解后视偏差，对比轨迹对 RL 区分生产性探索与冗余游荡
AgentRewind: Recoverable Execution for Long-Horizon LLM Agents
长程执行出错后缺恢复能力：AgentRewind 记录 agent 上下文与环境对齐 checkpoint，可回到早期状态带前次尝试信息续跑
Demystifying Agent Skills: Why They Work-Until They Don't
技能有效靠「程序锚定」稳定执行（占 65.7% 而非注入知识 4.5%）；检索是独立瓶颈，池 5→100 时实际使用精度 29.6%→3.3%
Handover of In-Context Learning State Across Session Boundaries
把跨会话交接形式化为「任务相关 ICL 状态传递」：预测等价性刻画最粗充分交接 + 固定 bit 需求，提出三段式交接记录
Agentic Transaction: Towards ACID-Compliant Agent Systems
把经典 ACID 事务语义重解释为 agent 执行的四个语义保障，用探索-执行-验证循环落地，广泛 benchmark 比含 Claude Code 在内的 SOTA 高 10.6%
Second Thought: Reasoning in Parallel as LLM Agents Act and Observe
ReAct 行动/观测间隙是推理空闲窗口：fork 四条辅助分支与主循环并行解码，把附加推理移出主线程串行路径，turn 数全降、Pass@1 不损
Practice Makes Unsafe: Skill Misevolution
精读 · 技能演化安全 · 不安全成功变持久策略
Seeker: Attention from Action
精读 · VLA · 从动作学注意力视觉瓶颈
FIRE-VLA: Failure-Informed Self-Evolution
精读 · VLA · 失败知情自进化驾驶
ContactGuard: Pre-Contact Execution Monitoring
精读 · 世界模型 · 接触前执行监控
The Objective Is the Bottleneck
精读 · 世界模型 · latent planner 目标函数瓶颈
ACPC: Diagnosing JEPA World Models
精读 · 世界模型 · JEPA 鲁棒性诊断
I-SDPO: Instance-Level Adaptive Self-Distillation
精读 · RL · 按实例路由蒸馏 vs GRPO
CrEST: Verifier-Bounded Credit Assignment
精读 · RL · CrEST verifier 天花板内的稠密信用分配
Temporal GRPO: Stage-Level Credit in VLA RL
精读 · RL · 阶段级优势替换轨迹级信用分配
Agent Behavioral Contracts II: Compositional Reliability
精读 · 多智能体 · 独立假设失效，有限样本无假设证书
E2-Explainer: Causal Communication Topologies
精读 · 多智能体 · 通信拓扑的因果归因解释
StateBridge: Training-free Latent Communication
精读 · 多智能体 · 隐状态对齐的无训练潜在通信
ERSkill: Skill-Guided Adaptive Memory Retrieval
精读 · Agent 记忆 · 检索机制作为可进化技能
RippleMem: Associative Recollection
精读 · Agent 记忆 · 线索锚点+关联扩张的联想回想
LycheeMemory V2: Segment-Level Consolidation
精读 · Agent 记忆 · 片段级批处理合并，构造 token 省 86%
ReFind: Raw Chat Logs Rival Structured Memory
精读 · Agent 记忆检索 · 零结构化预处理胜过图/树记忆
AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses
测试时强到弱能力迁移：builder构建推理harness，弱模型无参数更新ToM性能0.49→0.91，增益来自卸载不稳定推理进确定性代码
A Framework for Designing Reward Functions: From Objectives to Features to Human-Aligned Reward Functions
面向非专家的奖励函数设计流程：目标→结果变量→因果DAG子集→偏好elicitation，保持无冲突可行权重域
Retry, Switch, or Abstain? Learning Strategy-Aware Tool-Use Policies via Controlled Error Injection
BENCH2ROBUST把无故障工具基准转成受控随机环境，BTM无重训提升16.8pp，与RL结合达40.8-45.5%
Agent Skills Can Be Harmful: An Empirical Study of Skill-Induced Failures in LLM Agents
技能诱导失败差分归因：看似相关技能反导致错误实现或遗漏，Excessive Procedure源于过度验证与重型实现管线
The Sleeping Agent: What Gist-Based Context Compression Loses and Why
gist压缩保留关系/事件结构却丢弃日期时间，时间问题显著变差；一行提示修改使时间表达保留率提升20倍
Diagnosis Before Recovery: Turning Agent Failures into Selective Self-Correction
先诊断失败类型再裁剪恢复干预的DARC harness，让自修正从提示扩展变为恢复接口设计，ALFWorld/AppWorld/XBRL均提升
G0.5: One Autoregressive Stream for Robot Reasoning and Action
单流自回归VLA：推理与动作共享一套权重，R1真机76.7% vs π0.5的53.3%，VLM从上下文编码器变回决策者
Consolidator: Learning Persistent Routed Memory Across Context Boundaries
学习式记忆巩固：Consolidator把路由STM变换累积进LTM并反哺路由，仅12.35K参数提升更新映射召回至87.02%
XBridge: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication
词典锚点映射+潜在增强桥接解决异构LLM通信的实体grounding难题，7任务全胜文本通信且延迟低11×
GCPO: Diagnosing and Constraining Subspace Geometry in Rollout RL for LLMs
Principal-Subspace Overlap诊断rollout RL几何漂移，硬双边正交投影约束更新到互补子空间，超越GRPO/DAPO/GSPO
Beyond Memory: A Transactional Continuity Kernel for Long-Lived AI Agents
长生命周期Agent状态治理：激活契约解耦候选评估与原子激活，280万可达状态零不变量违反
Reinforcing Step-level Reasoning for Effective Self-Correction in LLMs
SFS-DPO两步RL框架：先强化步骤级推理，再显式训练自验证自修正，教师辅助变体注入错误验证rationale
Keep the Future, Drop the Rollout: RIFT for World Action Models
WAM延迟瓶颈：anticipation token一次前向构造完整未来K/V缓存，行动chunk延迟降68-89%，LIBERO 98.8%
The Devil Is in the Interface: Evaluating How Tool Architecture Shapes Coding Agent Behavior
工具架构（能力如何组织暴露）独立于能力集合改变编码Agent行为：CodeAct式接口省41.6%步骤与56.3% token
Self-Evolving Embodied Agents via Skill-Harness Evolution
冻结模型参数，通过目标环境rollout演化可复用技能与上下文代码harness，train-free具身适应
EvoGraph-Mem: Failure-Aware Editable Graph Memory for Long-Term Language Agents
可编辑见解图做失败感知的记忆维护，区分可复用与冲突/失效见解，跨骨干模型一致超越记忆基线
JEPA-WAM: 机器人世界模型的阶段级联合嵌入预测
JEPA-WAM: Stage-Level Joint-Embedding Prediction for World-Action Models in Robot Manipulation
检测到效应≠学会行动：LLM Acquisition Agent 的 Reward-SNR 下限
Detecting an Effect Is Not Learning to Act on It: A Reward-SNR Floor for LLM Acquisition Agents
TideRL: 就绪感知调度提升 Agentic RL Goodput
TideRL: Boosting Agentic RL Goodput with Readiness-Aware Scheduling
VibeLifeBench: 生活 agent 能否主动且持续？
VibeLifeBench: Can Your Life Agent Be Proactive and Persistent in a Living World?
ComBodied Agents: 以人为中心的 Agentic AI 新范式
ComBodied Agents: a New Paradigm of Human-Centric Agentic AI
ReTree: 树结构记忆的自我修正搜索 agent
Self-Correcting Long-Horizon Search Agents via Tree-Structured Memory
连续交互扩散：异步工具增强推理的 Diffusion-Native 运行时
Continuous Interaction Diffusion: A Diffusion-Native Runtime for Asynchronous Tool-Augmented Reasoning
不是铁板一块：中国前沿 LLM agent 的合作均衡差异
Not a Monolith: Lab-Level Divergence in the Cooperative Equilibria of Chinese Frontier LLM Agents
Agentic 系统的协同演化综述
Co-Evolution in Agentic Systems: Toward Self-Directed Evolution Beyond Human Design
为什么 CLAUDE.md 永远在涨？灾难性记忆
Why Does CLAUDE.md Keep Growing? Catastrophic Remembering in Agentic Coding
SkillZip: 免评估的结构化技能压缩
SkillZip: Evaluation-Free Skill Compression for Self-Evolving Agents by Discovering Reusable Structure
持久递归世界：让软件项目自主演化（EvoX Genesis）
Persistent Recursive Worlds Enable Autonomous Software Evolution
MAP-Graph: 多智能体工作流的溯源感知共享记忆
MAP-Graph: Provenance-Aware Shared Memory for Multi-Agent Workflows
从错误记忆到修正动作：依赖引导回滚修复
From Faulty Memories to Corrected Actions: Dependency-Guided Rollback Repair for Memory-Augmented Agents
Ex-Omni-2D: 表现力全模态对话与原生视觉呈现
Ex-Omni-2D: Expressive Omni-Modal Dialogue Models with Native Visual Presence
X2-Turn: 流式 ASR 与帧级轮转状态联合预测
X2-Turn: Frame-Synchronous Dual-Head Modeling for Joint Streaming ASR and Turn State Prediction
DuplexWorld: Can voice agents help you get through the day?
DuplexWorld: full-duplex voice agent benchmark — 6 worlds, 11 conversation types, turn-taking as independent metric. 用户指定精读，duplex 核心方向直接命中。
Social Gym: Verifiable Multi-Agent Social Skill Benchmark
Social Gym verifiable multi-agent social skill benchmark
MoRSE: Role-Subtask LoRA Experts for Multi-Agent LLM
MoRSE role-subtask specialized LoRA experts for multi-agent LLM
MasDrift: Authorization Drift in Multi-Agent Delegation
MasDrift authorization drift benchmark for multi-agent delegation
POLIS: Institutional Safety in Multi-Agent Systems
POLIS institutional governance safety for multi-agent systems
TAMS: Task-Aware Multi-View Adaptive Streaming for Teleoperation
TAMS task-aware multi-view adaptive streaming teleoperation
Agentic Harnesses: LLM-Judge Verification for Robot Planning
Agentic Harnesses LLM-judge verification layer for robot plans
Latent WM + Monotone Cost Ranking for Image-Goal Navigation
Latent world model with Monotone Cost Ranking for navigation
WorldSimProbe: Probing ACWM Simulator Fidelity
WorldSimProbe simulator fidelity evaluation for world models
Auto-Research as Greybox Fuzzing: Feedback-Driven Search
Auto-research as greybox fuzzing control loop
CTDE + HRA: Real-World Online RL for Robotic Manipulation
CTDE + HRA real-world online RL for robots
BCSD: Bidirectional Context Self-Distillation in RL
BCSD bidirectional context self-distillation for skill-based RL
CoRE: Consensus Rewards via Equilibrium for Test-Time RL
CoRE consensus rewards via equilibrium for test-time RL
LAVE: Latent Visual Evidence for Tool-Use in Video Agents
LAVE latent visual evidence for video tool agents
HyMeS: Memory-in-Code Hybrid Memory for VLA Manipulation
HyMeS hybrid memory for VLA robot manipulation
Muscle Memory: Beyond Retrieval for Personalizing Agents
Muscle Memory compilation for personalized agents
ReMEMBER: Missing-Evidence Memory for Streaming Dialogue Summarization
Streaming dialogue summarization — ReMEMBER missing-evidence memory framework
AtlasVLA: Persistent World-Ego State Modeling for Vision-Language-Action Models
持久世界状态记忆+自我工作状态记忆，把反应式 VLA 变成带状态追踪的长程操作
TEMPO: Semantic-Action Decoupled RL Post-Training for Vision-Language-Action Models
语义-动作解耦双时间尺度 TD3，冻结视觉主干分开优化语义级与动作级策略
Beyond Myopic World Models: Long-Horizon End-to-End Training
long-horizon end-to-end 直接预测训练，让梯度沿整条轨迹影响传播
CoBa: Cost-Effective Test-Time Scaling via Compute-Balanced Routing
TTS 建模为计算分配问题：下一个单位算力花在生成、验证还是停止
Every Cache Entry Earns Its Place: Global Allocation of Resolution and Coverage for KV Caches
可渐进精化原型树做 KV 全局分配，分辨率与覆盖率预算跨层/头自由流动
Trajectory-Relative Hindsight Distillation for Agentic Reinforcement Learning
轨迹相对 hindsight 蒸馏 + turn-aligned 打分协议，轮次间信用对齐分配
DiDPO: Diff-in-Diff Policy Optimization for Coding Agent Training
diff-in-diff 政策优化：把代码 agent 信用分配细化到 diff 级优势
How Much, Then Where: Credit-Conserving Action-to-Token Allocation for Multi-Turn Agent RL
两层信用分配：TD残差做动作信用，teacher-student似然差距做token分配
PsychoAgent: An Affect-Sensitive Cognitive Architecture for Language Agents
事实记忆+情感记忆双检索，冲突感知执行控制器整合，更类人的 agent 记忆
Agent Memory Distillation: Empowering Small LLM Agents with Memory from Large Language Models
训练-free 记忆蒸馏：把小模型 agent 注入 teacher 的结构化记忆
Explicit, Not Longer: What Makes Epistemic Stance Survive Memory Compression
记忆压缩后知识地位能否存活的关键是显式性而非长度
TEPA: Revoking Stale Memories for Conflict-Robust Language Agents
可证伪记忆：TEPA 把记忆有效性变成显式状态，trial-validated 撤销过期记忆
Interaction Creates Dynamical AI Behavior Absent in Isolation
非平衡统计物理解释 agent 交互涌现——交互是不可约的动力学源
BONSAI: Evolvability-Guided Tree Search over Skills
用 evolvability 引导技能树搜索，区分过拟合尖峰与宽阔平台
SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent
近端文本梯度下降优化 agent 技能，闭式前向更新+校验门控收缩
The Horizon Gap: Planning, Memory, Execution, Training, and Evaluation of Long-Horizon LLM Agents
Long-horizon agent 综述：规划/记忆/执行/训练/评测五维缺口（horizon gap）
EdgeXpert：面向MoE与投机解码的内存高效端侧LLM推理加速器
ω-0：面向并发人形移动操作的潜在预测世界动作模型
In-Context VLA：用上下文后训练与Agentic工具调用赋予VLA语言能力
SkillMemo：面向组合式具身操作的专家引导技能记忆框架
EvoHarness-RL：为长程LLM Agent学习自演化运行时harness
CIPO：面向搜索Agent的上下文信息策略优化
EnvACE：用世界预演内化环境动态的Agentic强化学习
AgentOPSD：面向Agentic强化学习的递归自蒸馏轮级信用分配
Project2Task：面向自主研究的图引导项目级规划
工具调用的苦涩教训
用Koopman谱分析认证多智能体系统的集体推理
多智能体LLM系统推理时并行的两层视角
视觉工具使用的幻觉：对"用图像思考"的因果审计
当历史在撒谎：误导性多轮历史下的工具调用评测与改进
OrchestraBench：多智能体编排故障模式与恢复能力评测
FormBharo：面向印度农村的对话式表单填写语音Agent
当策略改变概率：LLM代码审查中的模块化决策
别重生成，去调试：修复"近失败"硬件算子的领域专用Debug Agent
PRWeaver：对抗性恶意PR审计Benchmark
关系先验是收敛压力，不是涨点手段
CUADebug：计算机使用Agent的错误诊断与修复
并发Agent系统的有状态治理
多智能体LLM辩论中偏见共识的相变涌现
多智能体群体中的多元无知现象
MemArena：端侧个人记忆助手的第一人称大规模评测基准
Self-Evolving Coding Agents：给自我进化代码 Agent 画一张地图
重新审视 Agent 技能的自我进化
ADRS：自蒸馏奖励塑形构造稠密Token级信用信号
更短的推理，更早的答案？——推理接口评估
TraceCompiler：把Agent执行痕迹编译成确定性工作流
Verified Tool Calls — 非原子工具调用的验证层
Fail-Fast, Restart-Smart：给 SWE Agent 装一个轻量失败预警器
精读：DuplexSLA — 语音/语言/动作同步解码的原生全双工基础模型
补充精读：原生全双工 Speech-Language-Action 模型，160ms 共享时钟联合解码，语义驱动打断控制 + 会话中工具调用，与 DuplexOmni 架构对照
精读：DuplexChat — 大规模说话人分离全双工对话语料构建流水线
DuplexChat-Pipe 流水线从公开播客构建41.5万小时说话人分离全双工对话语料，规模超越Fisher/CallHome，但尚无下游SDLM训练验证
精读：全双工语音对话系统综述 — 架构层级 / 交互本体 / 决策状态机
L0-L3架构层级、T×I×R交互本体、决策状态机三框架审计全双工语音系统，揭示"实现差距"与"表面vs实质双工"
精读：Ring-Zero — 万亿参数规模的 Zero RL 涌现推理
四阶段训练流程 + clipped重要性采样 + 训练-推理ratio校正，把zero RL扩展到万亿参数规模，涌现5种认知行为
精读：DuplexPO — 用RL解耦全双工对话动态
DuplexPO 精读：FCDR + 动态关键窗口 + GRPO 的全双工 RL 训练，RL 扎实度评估，可借鉴点分析
精读：Kimi k1.5 — 产品级长CoT RL配方 + Long2Short
Moonshot AI技术报告精读。Length penalty原始公式、Long2Short四种方法（模型合并/最短RS/DPO/Long2Short RL）详解，Partial Rollouts基础设施，k1.5→DIET→f(g(x))谱系分析。
DIET 精读 — Difficulty-Aware Token Compression (NeurIPS 2025)
DIET：难度感知 token 压缩，GRPO Advantage Weighting 修复，同时提升准确率与降低 token 用量
Optima: Optimizing Effectiveness and Efficiency for LLM-Based Multi-Agent System
精读：MAS 通信效率可训练；三元奖励 + MCTS-DPO 迭代范式；DIET 与 f(g(x)) 的思想谱系
行为先于结果 — Agent训练数据质量的新范式
DLCoT/Recon/ASTER/WebSailor/Behavior Priming 五篇精读综合：五篇独立论文从不同角度收敛到同一结论——训练数据的行为质量决定上限，结果正确性只是必要非充分条件。
Behavior Priming: Beneficial Reasoning Behaviors in Agentic Search
CMU LTI：识别4类有益推理行为（信息核实、权威性评估、自适应搜索、错误恢复），SFT行为灌注后跑RL，比直接RL提升+37.2%。
WebSailor: Navigating Super-human Reasoning for Web Agent
Alibaba Tongyi：开源web agent与专有系统差距的根源是不确定性缩减能力，不是工具机制。DUPO算法 + 信息遮蔽训练数据，BrowseComp-zh 30.1%。
ASTER: Agentic Scaling with Tool-integrated Extended Reasoning
Interaction Collapse 问题与高密度 SFT 冷启动方案，4B 超越 235B — 华为+南大
Recon: Reconstruction-Guided Reasoning Synthesis
推理质量应用重建保真度衡量，而非后验合理化 — UC Berkeley
DLCoT: Deconstructing Long Chain-of-Thought
长 CoT 蒸馏质量的关键是主干内部的多样性，而非纯净度 — Alibaba+NYU
精读：Beyond the Leaderboard — LLM Agent 六类失败模式分类法
综合27篇论文、19个benchmark，提炼LLM agent跨领域失败模式统一分类法
精读：AgenticAI-Supervisor — 可扩展 Agentic RL 仿真环境
Uber AI 的 agentic RL 仿真平台：三维度奖励设计，纯结果奖励导致40%约束违反
精读 | UI-MOPD: 跨平台 GUI Agent 的多教师在线蒸馏
精读：UI-MOPD 多教师在线蒸馏解决跨平台 GUI agent 灾难性遗忘问题，OSWorld 38.2% / MobileWorld 12.0%
A-TMA: Ghost Memory & Three-Layer State-Aware Agent Memory
深读：A-TMA — Agent 长期记忆状态失效解耦（三层框架 + 内嵌论文图表）
AgenticSTS 深读
AgenticSTS 深读笔记 — 有界记忆合约测试床
Agentic Data Synthesis 精读 v2 — 结构化深度解析
四篇论文结构化精读：做了什么 / 评估设计 / 结果 vs 基线。含论文切换、结果对比多视图、图片放大、阅读模式等交互
Agentic Data Synthesis 精读 — 四篇论文深度解析
TMax / OpenThoughts-Agent / Autodata / AgentSkiller 四篇论文图文并茂深度精读