每日 AI / 电商 / 推荐系统 / Agent 论文精选https://slinene.github.io/https://slinene.github.io/paper/2026-11-03-arloop-dai-zhu-yi-li-can-chai-de-wan-quan-kai-yuan-yu-xun-261105234/https://slinene.github.io/paper/2026-11-03-arloop-dai-zhu-yi-li-can-chai-de-wan-quan-kai-yuan-yu-xun-261105234/提出 ARLoop，用滑动窗口深度注意力改进循环 Transformer，1.2B/0.5T 模型超过 Huginn-3.5B 并接近 Qwen3.5-2BTue, 03 Nov 2026 00:00:00 GMTTrainingLooped TransformersAttention ResidualsPretrainingSliding-Window AttentionParameter-Efficient LLMRecurrent Depthhttps://slinene.github.io/paper/2026-11-03-dualevolve-tui-jian-xi-tong-ji-zhi-yu-bian-pai-de-shuang-ji-261104439/https://slinene.github.io/paper/2026-11-03-dualevolve-tui-jian-xi-tong-ji-zhi-yu-bian-pai-de-shuang-ji-261104439/同时演化推荐模型代码和 Agent 编排，用过程质量反馈与 Thompson 采样在 64 个比较中全部优于现有自演化方法Tue, 03 Nov 2026 00:00:00 GMTRecSysSelf-EvolutionRecommender SystemsLLM AgentsThompson SamplingProcess RewardHarness Optimizationhttps://slinene.github.io/paper/2026-11-03-harp-mian-xiang-llm-tui-jian-de-xun-lian-free-shuang-hua-xin-261105319/https://slinene.github.io/paper/2026-11-03-harp-mian-xiang-llm-tui-jian-de-xun-lian-free-shuang-hua-xin-261105319/将 LLM 推荐重构为用户-物品双 Profile 通信，用检索侧同质性门控和反思更新提升推荐准确率Tue, 03 Nov 2026 00:00:00 GMTRecSysLLM RecommendationAgentic CommunicationDual ProfileHomophilyRetrievalTraining-Freehttps://slinene.github.io/paper/2026-11-03-intentgeo-mian-xiang-tui-jian-chang-jing-de-yi-tu-gan-zhi-261104225/https://slinene.github.io/paper/2026-11-03-intentgeo-mian-xiang-tui-jian-chang-jing-de-yi-tu-gan-zhi-261104225/提出 AMO，用固定推荐指令核心加学习追加策略，防止 prompt drift，提升多意图下 item 可见性Tue, 03 Nov 2026 00:00:00 GMTGenRecGenerative Engine OptimizationLLMRecommendationPrompt OptimizationIntent-awareCitation Rankhttps://slinene.github.io/paper/2026-11-03-jhima-webrl-chang-cheng-wang-ye-ren-wu-de-lian-he-fen-ceng-261105263/https://slinene.github.io/paper/2026-11-03-jhima-webrl-chang-cheng-wang-ye-ren-wu-de-lian-he-fen-ceng-261105263/教师生成子任务监督，分层GRPO联合训练7B规划器与7B执行器，WebArena-Lite-v2达37.2%，超32B RL基线13.2个百分点Tue, 03 Nov 2026 00:00:00 GMTMultiAgentHierarchical RLMulti-AgentWeb AgentGRPOTask DecompositionPlanninghttps://slinene.github.io/paper/2026-11-03-ni-bu-xu-yao-chang-shang-xia-wen-pin-fan-ya-suo-zhi-chi-duan-261105306/https://slinene.github.io/paper/2026-11-03-ni-bu-xu-yao-chang-shang-xia-wen-pin-fan-ya-suo-zhi-chi-duan-261105306/频繁上下文压缩与选择性读取，使长程/长输入任务无需长上下文，准确率基本持平且成本大降Tue, 03 Nov 2026 00:00:00 GMTAgentlong contextcontext compactionsemantic stateselective readingagentic LLMcost efficiencyhttps://slinene.github.io/paper/2026-11-03-onetrans-gsr-dan-zi-hui-gui-transformer-tong-yi-zhong-pai-de-261104720/https://slinene.github.io/paper/2026-11-03-onetrans-gsr-dan-zi-hui-gui-transformer-tong-yi-zhong-pai-de-261104720/单自回归 Transformer 统一重排/混排直接生成 slate，线上 GMV/u +2.63%、广告价值 +2.06%、复购 +2.03%，延迟降 13%Tue, 03 Nov 2026 00:00:00 GMTGenRecGenerative RecommendationSlate GenerationReinforcement LearningReward ModelRerankingMix-Rankinghttps://slinene.github.io/paper/2026-11-03-reusegr-zi-gua-ying-qian-zhui-fu-yong-jie-ou-sheng-cheng-shi-261104259/https://slinene.github.io/paper/2026-11-03-reusegr-zi-gua-ying-qian-zhui-fu-yong-jie-ou-sheng-cheng-shi-261104259/7B离线规划共享SID前缀，0.5B在线个性化补全，以38.9%更低延迟达到与7B相当HR@100Tue, 03 Nov 2026 00:00:00 GMTGenRecGenerative RetrievalSemantic IDPrefix CachingKnowledge DistillationSearch AdvertisingOffline-Online Decouplinghttps://slinene.github.io/paper/2026-11-03-shu-ju-shou-xian-chou-mi-jian-suo-de-suo-fang-lu-yu-suan-li-261104979/https://slinene.github.io/paper/2026-11-03-shu-ju-shou-xian-chou-mi-jian-suo-de-suo-fang-lu-yu-suan-li-261104979/提出 MAEL 作为数据受限微调的稳定目标，拟合乘法 scaling law 并量化模型规模相对数据规模的计算效率优势Tue, 03 Nov 2026 00:00:00 GMTTrainingscaling lawsdense retrievaldata-constrained trainingcompute efficiencyfine-tuninghttps://slinene.github.io/paper/2026-11-03-spade-mian-xiang-ben-di-sou-suo-de-shi-kong-pian-hao-dui-qi-261104662/https://slinene.github.io/paper/2026-11-03-spade-mian-xiang-ben-di-sou-suo-de-shi-kong-pian-hao-dui-qi-261104662/联合优化词表、时空编码、DPO 对齐与解码多样性，显著提升本地生活 query 推荐Tue, 03 Nov 2026 00:00:00 GMTQueryRecQuery RecommendationLocal SearchGenerative RecommendationDPOSpatiotemporal EncodingDiversity Decodinghttps://slinene.github.io/paper/2026-11-03-t-rec-books-kong-zhi-xu-qiu-pi-lu-de-zhi-neng-tui-jian-xi-xu-261104459/https://slinene.github.io/paper/2026-11-03-t-rec-books-kong-zhi-xu-qiu-pi-lu-de-zhi-neng-tui-jian-xi-xu-261104459/提出控制需求披露时机的书籍推荐评测基准，揭示隐藏需求下智能体成功率从 0.975 跌至 0.283Tue, 03 Nov 2026 00:00:00 GMTEvalAgentic Recommender SystemsRequirement ElicitationBenchmarkUser SimulationStructured SearchLLM Agentshttps://slinene.github.io/paper/2026-11-03-you-yi-yi-de-id-tui-jian-xi-tong-yu-yi-idshi-cao-jiao-cheng-261104140/https://slinene.github.io/paper/2026-11-03-you-yi-yi-de-id-tui-jian-xi-tong-yu-yi-idshi-cao-jiao-cheng-261104140/系统梳理从原始ID到语义ID的演进，覆盖生成式与判别式应用并提供可动手的实践教程Tue, 03 Nov 2026 00:00:00 GMTGenRecSemantic IDGenerative RecommendationRQ-VAECross-DomainTutorialMultimodalhttps://slinene.github.io/paper/2026-11-03-zerogr-sheng-cheng-shi-tui-jian-de-guo-lu-you-xian-yue-shu-261104715/https://slinene.github.io/paper/2026-11-03-zerogr-sheng-cheng-shi-tui-jian-de-guo-lu-you-xian-yue-shu-261104715/提出过滤优先的 GPU 约束解码系统，先滤除非法 SID 再算 logits，召回与先打分几乎一致且延迟最高降 9.9×Tue, 03 Nov 2026 00:00:00 GMTGenRecConstrained DecodingGenerative RecommendationSemantic IDGPU ServingBeam SearchFiltering Indexhttps://slinene.github.io/paper/2026-11-02-cai-jue-xue-xi-yan-zheng-qi-yin-dao-de-duo-zhi-neng-ti-dian-261102758/https://slinene.github.io/paper/2026-11-02-cai-jue-xue-xi-yan-zheng-qi-yin-dao-de-duo-zhi-neng-ti-dian-261102758/提出 CADRE 框架，用冻结三专家加可训练 Verifier 做选择性辩论，并以成功经验回放 GRPO 优化后蒸馏到轻量学生模型Mon, 02 Nov 2026 00:00:00 GMTRecSysMulti-AgentRelevance ModelingGRPOExperience ReplayKnowledge DistillationE-Commerce Searchhttps://slinene.github.io/paper/2026-11-02-coworkos-mian-xiang-xian-shi-zi-dong-hua-de-ke-kuo-zhan-zhi-261103178/https://slinene.github.io/paper/2026-11-02-coworkos-mian-xiang-xian-shi-zi-dong-hua-de-ke-kuo-zhan-zhi-261103178/提出三层架构（Main、Coordinator、Subagents）的通用长时程智能体运行时，支持人机共享工件与持久状态，多个基准上超越主流框架Mon, 02 Nov 2026 00:00:00 GMTMultiAgentAgent HarnessMulti-AgentLong-horizonHuman-in-the-loopPersistent MemoryRSIhttps://slinene.github.io/paper/2026-11-02-cua-master-kai-fang-shi-jie-di-gui-gai-jin-qu-dong-de-ji-ji-261103126/https://slinene.github.io/paper/2026-11-02-cua-master-kai-fang-shi-jie-di-gui-gai-jin-qu-dong-de-ji-ji-261103126/提出嵌套双循环递归改进框架 CUA-Master，结合开放世界数据与自催化传播，显著提升 GUI 智能体性能。Mon, 02 Nov 2026 00:00:00 GMTAgentCUAData FlywheelGUI AgentsTrajectory ReconstructionSelf-ImprovementTraininghttps://slinene.github.io/paper/2026-11-02-dian-shang-gou-wu-agent-de-sheng-cheng-shi-jiang-li-mo-xing-261103008/https://slinene.github.io/paper/2026-11-02-dian-shang-gou-wu-agent-de-sheng-cheng-shi-jiang-li-mo-xing-261103008/从真实购物行为回溯构造评价链，三阶段训练生成式奖励模型，在线提升淘宝购物 Agent 有效协助率与满意度Mon, 02 Nov 2026 00:00:00 GMTAgentGenerative Reward ModelE-Commerce AgentRLVRGRPOEvaluation ChainPersonalizationhttps://slinene.github.io/paper/2026-11-02-dr-pixel-shi-jue-yu-yan-mo-xing-qu-dong-de-ye-mian-yuan-shen-261103097/https://slinene.github.io/paper/2026-11-02-dr-pixel-shi-jue-yu-yan-mo-xing-qu-dong-de-ye-mian-yuan-shen-261103097/首个开源端到端多模态深度研究代理，直接检索渲染页面图像并引入反事实信用重加权训练，超越 OCR 基线Mon, 02 Nov 2026 00:00:00 GMTAgentDeep ResearchVision-Language ModelsPage-Native RetrievalCounterfactual Credit ReweightingAgent Traininghttps://slinene.github.io/paper/2026-11-02-duo-yang-jiao-shi-zai-xian-ce-lue-zheng-liu-yong-yu-sheng-261103272/https://slinene.github.io/paper/2026-11-02-duo-yang-jiao-shi-zai-xian-ce-lue-zheng-liu-yong-yu-sheng-261103272/提出 HERO，通过异构教师路由和多样性物品 rollout 缓解生成式推荐中 on-policy distillation 的流行度偏差Mon, 02 Nov 2026 00:00:00 GMTGenRecGenerative RecommendationOn-Policy DistillationPopularity BiasSemantic IDKnowledge DistillationDiversityhttps://slinene.github.io/paper/2026-11-02-ji-yu-cai-xiang-yu-fan-bo-de-llm-agent-dong-zuo-yu-zhi-xing-261102950/https://slinene.github.io/paper/2026-11-02-ji-yu-cai-xiang-yu-fan-bo-de-llm-agent-dong-zuo-yu-zhi-xing-261102950/提出 BLACKSWAN：对每个会改变环境的动作，先预测效果、再生成可测试的怀疑并验证，执行前主动寻找反驳证据Mon, 02 Nov 2026 00:00:00 GMTAgentLLM AgentAction VerificationRefutationPre-execution VerificationAppWorldSelf-correctionhttps://slinene.github.io/paper/2026-11-02-lu-you-ren-yi-tong-yong-aixi-tong-mo-xing-yu-ce-lue-lu-you-261105275/https://slinene.github.io/paper/2026-11-02-lu-you-ren-yi-tong-yong-aixi-tong-mo-xing-yu-ce-lue-lu-you-261105275/构建覆盖25个基准374k轨迹的omni-modal路由框架，证明学习路由器超最佳固定候选并提升效率8.82倍Mon, 02 Nov 2026 00:00:00 GMTAgentmodel routingmulti-modalquality-cost trade-offlearned routergeneral-purpose AIhttps://slinene.github.io/paper/2026-11-02-mei-ge-zi-ju-du-suan-shu-jiang-zhi-ling-chai-cheng-ke-hui-da-261103179/https://slinene.github.io/paper/2026-11-02-mei-ge-zi-ju-du-suan-shu-jiang-zhi-ling-chai-cheng-ke-hui-da-261103179/把带指令查询拆为主题与多个类型化子句，逐项用小型 judge 输出概率并相乘，零样本提升指令遵循重排效果Mon, 02 Nov 2026 00:00:00 GMTRecSysinstruction-followingrerankingzero-shotLLM judgesearchprobabilistic scoringhttps://slinene.github.io/paper/2026-11-02-mian-xiang-llm-tui-jian-de-duo-jiao-shi-tui-li-zheng-liu-yu-261102457/https://slinene.github.io/paper/2026-11-02-mian-xiang-llm-tui-jian-de-duo-jiao-shi-tui-li-zheng-liu-yu-261102457/MUSE 构建多教师+学生自探索候选池，用 Rank-Gain 效用筛选并对比蒸馏，让 LLM 推荐器学到真正有益的偏好推理Mon, 02 Nov 2026 00:00:00 GMTRecSysReasoning DistillationLLM4RecMulti-TeacherContrastive LearningSelf-ExplorationRank-Gain Utilityhttps://slinene.github.io/paper/2026-11-02-pacerec-ji-yu-xie-tong-zheng-ju-de-llm-wu-pin-qian-ru-jian-261103068/https://slinene.github.io/paper/2026-11-02-pacerec-ji-yu-xie-tong-zheng-ju-de-llm-wu-pin-qian-ru-jian-261103068/用 flow matching 学共享速度场，将静态 LLM 物品嵌入渐进适配前序行为目标，实现无 ID 长尾序列推荐Mon, 02 Nov 2026 00:00:00 GMTRecSysSequential RecommendationLLM EmbeddingsFlow MatchingLong-tailID-freeItem Representationhttps://slinene.github.io/paper/2026-11-02-recrsi-ji-zhi-yin-dao-de-tui-jian-suan-fa-zi-dong-fa-xian-261102487/https://slinene.github.io/paper/2026-11-02-recrsi-ji-zhi-yin-dao-de-tui-jian-suan-fa-zi-dong-fa-xian-261102487/用机制假设组织推荐算法搜索，在16个设置上平均NDCG@5相对初始模型提升8.48%Mon, 02 Nov 2026 00:00:00 GMTRecSysAutomated Algorithm DiscoveryRecommendation SystemsLLM AgentsMechanism HypothesisExperiment FeedbackSearch Efficiencyhttps://slinene.github.io/paper/2026-11-02-shi-pin-sheng-cheng-bei-hou-kling-duo-mo-tai-li-jie-de-guan-261104916/https://slinene.github.io/paper/2026-11-02-shi-pin-sheng-cheng-bei-hou-kling-duo-mo-tai-li-jie-de-guan-261104916/从数据-条件-对齐-评估链剖析多模态理解如何提升可控视频生成质量Mon, 02 Nov 2026 00:00:00 GMTMultimodalMultimodal UnderstandingVideo GenerationCaptioningPrompt EnhancementReward ModelEvaluationhttps://slinene.github.io/paper/2026-11-02-shq-mian-xiang-llm-quan-zhong-liang-hua-de-ke-xue-xi-wang-ge-261102966/https://slinene.github.io/paper/2026-11-02-shq-mian-xiang-llm-quan-zhong-liang-hua-de-ke-xue-xi-wang-ge-261102966/用两参数 sinh 压缩网格替代固定查找表，端到端 KL 搜索参数，3-4bit 量化误差显著低于 NF/BOF/INTMon, 02 Nov 2026 00:00:00 GMTOtherLLM quantizationlookup-tableblock-scaledcompanding gridKL divergenceGPTQhttps://slinene.github.io/paper/2026-11-02-strategist-yong-yuan-zhi-neng-ti-jin-hua-tui-li-ce-lue-bing-261103618/https://slinene.github.io/paper/2026-11-02-strategist-yong-yuan-zhi-neng-ti-jin-hua-tui-li-ce-lue-bing-261103618/将推理策略写成可复用蓝图并跨任务搜索积累，在 30 个基准上以 39% 更低成本匹配最强基线Mon, 02 Nov 2026 00:00:00 GMTAgentMeta-AgentInference StrategyBlueprint LibraryMulti-tier RoutingCross-task Transferhttps://slinene.github.io/paper/2026-11-02-tong-guo-yin-shi-pin-pai-zhu-ru-ke-hua-dui-kang-xing-sheng-261104005/https://slinene.github.io/paper/2026-11-02-tong-guo-yin-shi-pin-pai-zhu-ru-ke-hua-dui-kang-xing-sheng-261104005/提出多代理框架IBI，通过单句隐式品牌注入提升生成式搜索中的品牌可见度，并构建IBIBench基准Mon, 02 Nov 2026 00:00:00 GMTRAGAdversarial GEOBrand InjectionMulti-AgentRAGBenchmarkLLM Searchhttps://slinene.github.io/paper/2026-11-02-toolspace-yi-gong-ju-wei-zhong-xin-de-chang-cheng-zhi-neng-261105036/https://slinene.github.io/paper/2026-11-02-toolspace-yi-gong-ju-wei-zhong-xin-de-chang-cheng-zhi-neng-261105036/用真实执行轨迹构建加权工具组合图，按需合成环境并生成可回放验证的长程跨域任务Mon, 02 Nov 2026 00:00:00 GMTAgentTool Composition GraphEnvironment SynthesisLong-Horizon AgentsMCPSFT-RLReplay Verificationhttps://slinene.github.io/paper/2026-11-02-wei-diao-shu-ju-yu-chu-li-xuan-ze-dui-llmtui-jian-xiao-yong-261103113/https://slinene.github.io/paper/2026-11-02-wei-diao-shu-ju-yu-chu-li-xuan-ze-dui-llmtui-jian-xiao-yong-261103113/系统评估微调预处理对LLM生成式推荐效用与公平性的影响，并验证目录约束解码与提示增强的有效性Mon, 02 Nov 2026 00:00:00 GMTGenRecLLM4RecFine-tuningConstrained DecodingFairnessData PreprocessingPrompt Augmentationhttps://slinene.github.io/paper/2026-11-02-xiao-yong-gan-zhi-sheng-cheng-shi-tui-jian-rong-he-xiao-yong-261102981/https://slinene.github.io/paper/2026-11-02-xiao-yong-gan-zhi-sheng-cheng-shi-tui-jian-rong-he-xiao-yong-261102981/将多目标效用写入SID末级并引导beam search，降低高价值候选被剪枝Mon, 02 Nov 2026 00:00:00 GMTGenRecGenerative RecommendationSemantic IDUtility-Aware DecodingBeam SearchMulti-ObjectiveLookup-Free Quantizationhttps://slinene.github.io/paper/2026-11-02-xue-sheng-qian-zhui-jiao-shi-gua-ying-rang-jiao-shi-gua-ying-261103133/https://slinene.github.io/paper/2026-11-02-xue-sheng-qian-zhui-jiao-shi-gua-ying-rang-jiao-shi-gua-ying-261103133/在蒸馏前用学生前缀 RLVR 微调教师，提升其对学生在策略轨迹的监督可靠性，无需改变蒸馏目标Mon, 02 Nov 2026 00:00:00 GMTTrainingOn-Policy DistillationTeacher AdaptationRLVRKnowledge DistillationPrefix Conditioninghttps://slinene.github.io/paper/2026-11-01-axion-1-0-jian-hua-de-kai-fang-zhi-neng-ti-ji-chu-mo-xing-261103233/https://slinene.github.io/paper/2026-11-01-axion-1-0-jian-hua-de-kai-fang-zhi-neng-ti-ji-chu-mo-xing-261103233/用两阶段 SFT 加 RL 的极简 recipe，构建可复现的开放 Agent 基础模型，在多项 agentic 基准上达到前沿水平Sun, 01 Nov 2026 00:00:00 GMTAgentAgentRLSFTSynthetic DataTool UseGRPOhttps://slinene.github.io/paper/2026-11-01-credo-mian-xiang-shi-xu-tui-li-zhi-neng-ti-de-zheng-ju-tiao-261101643/https://slinene.github.io/paper/2026-11-01-credo-mian-xiang-shi-xu-tui-li-zhi-neng-ti-de-zheng-ju-tiao-261101643/CREDO 用中间证据恢复局部监督，改进 GRPO 在时序知识图谱问答中的信用分配，提升 MultiTQ/TimelineKGQA 准确率Sun, 01 Nov 2026 00:00:00 GMTTrainingTKGQAGRPOCredit AssignmentRLHFLLM AgentsTemporal Reasoninghttps://slinene.github.io/paper/2026-11-01-evorec-tong-yi-zi-jin-hua-zhi-neng-ti-tui-jian-yu-dong-tai-261101977/https://slinene.github.io/paper/2026-11-01-evorec-tong-yi-zi-jin-hua-zhi-neng-ti-tui-jian-yu-dong-tai-261101977/提出统一 agentic 推荐框架，通过动态反思树定位失败并生成可复用修复策略，实现无参数更新的自进化推荐Sun, 01 Nov 2026 00:00:00 GMTAgentAgentic RecommendationDynamic Reflection TreeSelf-Evolving RecommenderLLM PlanningIndustrial A-B Testhttps://slinene.github.io/paper/2026-11-01-leanguard-mian-xiang-gong-ju-diao-yong-zhi-neng-ti-de-lean-261102332/https://slinene.github.io/paper/2026-11-01-leanguard-mian-xiang-gong-ju-diao-yong-zhi-neng-ti-de-lean-261102332/在 LLM 智能体与工具之间嵌入 Lean 4 策略运行时，以形式化证明保证每次工具调用合规，并支持反馈或证明引导修复Sun, 01 Nov 2026 00:00:00 GMTAgentLLM AgentsFormal VerificationRuntime GuardrailsPrompt InjectionTool CallingLean 4https://slinene.github.io/paper/2026-11-01-recursa-mai-xiang-di-gui-zi-gai-jin-de-ji-chu-mo-xing-xi-261104035/https://slinene.github.io/paper/2026-11-01-recursa-mai-xiang-di-gui-zi-gai-jin-de-ji-chu-mo-xing-xi-261104035/构建知识仓库引导的多智能体递归自改进系统，无更强模型下将法律领域任务全对率最高提升至 55.0%Sun, 01 Nov 2026 00:00:00 GMTMultiAgentRecursive Self-ImprovementMulti-AgentKnowledge RepositoryHarnessDomain AdaptationLong-Horizon Agentshttps://slinene.github.io/paper/2026-11-01-sheng-cheng-shi-ai-sou-suo-sui-ji-shi-yan-zhong-de-yong-hu-261101582/https://slinene.github.io/paper/2026-11-01-sheng-cheng-shi-ai-sou-suo-sui-ji-shi-yan-zhong-de-yong-hu-261101582/提出 Session Success Rate (SSR)，用人类标注会话训练可解释模型区分点击与放弃中的成功/失败，在线实验灵敏度约为 Sessions per User 的 2.5 倍Sun, 01 Nov 2026 00:00:00 GMTEvalGood AbandonmentSearch MetricsRandomized ExperimentsLLM SearchUser SuccessOnline Evaluationhttps://slinene.github.io/paper/2026-11-01-specpilot-tou-ji-jie-ma-de-zi-zhu-suan-fa-fa-xian-yu-xi-tong-261102501/https://slinene.github.io/paper/2026-11-01-specpilot-tou-ji-jie-ma-de-zi-zhu-suan-fa-fa-xian-yu-xi-tong-261102501/用自治 Agent 循环自动发现投机解码的算法与系统优化，单请求加速比提升 2.4 倍、批处理提升 2 倍Sun, 01 Nov 2026 00:00:00 GMTLLMSpeculative DecodingAgentic LoopLLM InferenceCost ModelReward HackingSystems Optimizationhttps://slinene.github.io/paper/2026-11-01-xiang-tong-gui-ji-bu-tong-zhi-shi-mian-xiang-sou-suo-agent-261101987/https://slinene.github.io/paper/2026-11-01-xiang-tong-gui-ji-bu-tong-zhi-shi-mian-xiang-sou-suo-agent-261101987/提出 ISA-Distill，按学生状态的事实可恢复性加权 fact-use 与 evidence-seeking 双视图反馈，纠正 teacher 特权信息偏差Sun, 01 Nov 2026 00:00:00 GMTAgentagentic searchon-policy distillationprocess rewardfactual recoverabilitymulti-hop QAcredit assignmenthttps://slinene.github.io/paper/2026-11-01-zheng-ju-yin-dao-de-zi-gua-ying-tu-jian-suo-eagr-261101910/https://slinene.github.io/paper/2026-11-01-zheng-ju-yin-dao-de-zi-gua-ying-tu-jian-suo-eagr-261101910/用 VectorDB 元数据中的语义描述符构成隐式图，以积累证据动态选择检索方向，无需显式图或查询改写Sun, 01 Nov 2026 00:00:00 GMTRAGRAGMulti-hop QAGraph RetrievalEvidence-GuidedVectorDB MetadataAdaptive Retrievalhttps://slinene.github.io/paper/2026-10-31-northcua-tong-guo-da-gui-mo-jing-yan-fei-lun-jia-su-ji-suan-261102102/https://slinene.github.io/paper/2026-10-31-northcua-tong-guo-da-gui-mo-jing-yan-fei-lun-jia-su-ji-suan-261102102/用经验飞轮、分级学习与桌面-网页联合强化学习，训练出开源SOTA的计算机使用代理模型族Sat, 31 Oct 2026 00:00:00 GMTAgentComputer-Use AgentExperience FlywheelReinforcement LearningGUI GroundingStaged LearningOpen-sourcehttps://slinene.github.io/paper/2026-10-31-paretosaddler-mian-xiang-zhi-liang-yu-cheng-ben-de-zhi-neng-261102079/https://slinene.github.io/paper/2026-10-31-paretosaddler-mian-xiang-zhi-liang-yu-cheng-ben-de-zhi-neng-261102079/分离中间候选提升与最终部署选择，用 Bayesian bootstrap Pareto 支持联合优化 Agent 质量与成本Sat, 31 Oct 2026 00:00:00 GMTAgentAgent HarnessPareto OptimizationBayesian BootstrapQuality-Cost TradeoffLLM Agentshttps://slinene.github.io/paper/2026-10-31-swe-trail-li-shi-yin-dao-de-bian-cheng-agent-gui-ji-cai-ji-261102074/https://slinene.github.io/paper/2026-10-31-swe-trail-li-shi-yin-dao-de-bian-cheng-agent-gui-ji-cai-ji-261102074/用历史轨迹引导采集资源分配但不进入 agent 上下文，训练收益提升 2.33/1.60 ppSat, 31 Oct 2026 00:00:00 GMTTrainingtrajectory collectioncoding agentpost-trainingSWE-benchhistory-guidedtrain-test mismatchhttps://slinene.github.io/paper/2026-10-31-xu-shu-zai-ce-lue-zheng-liu-zhi-xue-jiao-shi-top-k-pai-xu-261102112/https://slinene.github.io/paper/2026-10-31-xu-shu-zai-ce-lue-zheng-liu-zhi-xue-jiao-shi-top-k-pai-xu-261102112/O3PD 在策略蒸馏中只复现教师 top-K token 排序，不拟合概率值，对温度不变并在数学推理上全面优于 OPDSat, 31 Oct 2026 00:00:00 GMTTrainingKnowledge DistillationOn-Policy DistillationRanking LossLLMTemperature Invarianthttps://slinene.github.io/paper/2026-10-30-browseflow-yong-yu-sheng-cheng-shi-yong-hu-mo-ni-de-xi-shu-261041374/https://slinene.github.io/paper/2026-10-30-browseflow-yong-yu-sheng-cheng-shi-yong-hu-mo-ni-de-xi-shu-261041374/用离散 Flow Matching 对稀疏关键事件和浏览深度建模，生成随机用户浏览轨迹，在三个数据集上取得最优分布拟合Fri, 30 Oct 2026 00:00:00 GMTRecSysUser SimulationDiscrete Flow MatchingSparse EventsDistributional EvaluationRecommender SystemsTaobaoBrowsehttps://slinene.github.io/paper/2026-10-30-chronos-2-duo-bian-liang-shi-jian-xu-lie-yu-ce-ji-chu-mo-261041956/https://slinene.github.io/paper/2026-10-30-chronos-2-duo-bian-liang-shi-jian-xu-lie-yu-ce-ji-chu-mo-261041956/零样本多变量时序基础模型，联合建模序列交互与异构协变量，精度与推理效率显著领先Fri, 30 Oct 2026 00:00:00 GMTOthertime series foundation modelmultivariate forecastingzero-shotprobabilistic forecastingcovariatesChronos-2https://slinene.github.io/paper/2026-10-30-cong-chu-xing-xing-wei-wa-jue-yin-shi-yi-tu-de-kua-yu-poi-261041319/https://slinene.github.io/paper/2026-10-30-cong-chu-xing-xing-wei-wa-jue-yin-shi-yi-tu-de-kua-yu-poi-261041319/用 LLM 将地图出行行为与画像推理为有序自然语言 query，再用 GRPO 双层 credit assignment 对齐后续服务交互Fri, 30 Oct 2026 00:00:00 GMTQueryRecLLMGRPOCross-Domain POI RecommendationIntent InferenceQuery GenerationMobility Behaviorhttps://slinene.github.io/paper/2026-10-30-cros-ji-yu-da-luan-si-wei-lian-shun-xu-de-yin-guo-jiang-li-261042054/https://slinene.github.io/paper/2026-10-30-cros-ji-yu-da-luan-si-wei-lian-shun-xu-de-yin-guo-jiang-li-261042054/用打乱推理步骤顺序而非改变信息内容的因果奖励，缓解结果奖励无法保证 CoT 因果贡献的问题Fri, 30 Oct 2026 00:00:00 GMTTrainingRLVRCoT CausalityReasoning FaithfulnessReward DesignLLM Training