Latest AI News & Industry Updates

Papers Builders09:38

Chinese contentMLIPS：跨层共享索引池，vLLM 实测长上下文最高提速 1.40 倍

这篇论文教你让稀疏注意力模型多层共用一套候选池，不用重训，vLLM 里给 DeepSeek-V4-Flash 实测提速 1.4 倍还省电，做推理优化的值得看。

#MLIPS #DeepSeek-V4-Flash #vLLM #稀疏注意力

arXiv: DeepSeek@Samkit Jain et al. (3)Source

Papers Builders11:48

Chinese content研究：GEPA 与 SkillOpt 在 DeepSeek V4 Flash 和 Qwen3.8-27B 本地部署上的效果差异

GEPA 在 MATH-500 上能从 0.828 拉到 0.904，但在 HotpotQA 上完全没用，SkillOpt 还会翻车——如果你在本地跑推理模型做提示词优化，这篇能帮你避坑。

#GEPA #SkillOpt #DeepSeek #Qwen

Event

arXiv: DeepSeek@Andrea Giummarra et al. (6)2 sourcesSource

Papers BuildersSelected11:47

Addressed-Memory Writes Fail at Reasoning in Qwen3 and DeepSeek Beds

Memory write-injection looks like a cheap way to add knowledge to a deployed model, but these results show the added facts cannot be chained into multi-hop reasoning by the model itself — orchestration must live in the application layer. Builders wiring RAG-style memory into agents should chain written atoms explicitly in prompts or system logic, and benchmark vendors should adopt coverage, synergy and distortion reporting. The tested beds include Chinese frontier models Qwen3 and DeepSeek, making the findings broadly relevant across model families.

#memory #DeepSeek #Qwen #RAG

arXiv: DeepSeek@Wei Gao et al. (4)Source

Papers Builders11:47

Chinese content用 2022 EU Loneliness Survey 检验 LLM 能否复现孤独感年龄差异

拿四个 LLM 模拟欧洲四国调查，看谁能复现真实的年龄差距，结果三个模型翻车，适合做问卷模拟的人看看。

#LLM #DeepSeek #GPT-5.6 Terra #Mistral

arXiv: DeepSeek@Yuliya Li et al. (3)Source

Papers Builders11:46

Chinese contentCoEvo-RSI 论文提出模型与执行框架协同进化框架，缓解脑身不匹配问题

一篇研究智能体自我改进的论文，Qwen3.5-9B 用它这套框架跑赢了 DeepSeek-V4-Pro 和 GLM-5.3，做 agent 的可以看看思路。

#CoEvo-RSI #Qwen3.5-9B #智能体 #递归自我改进

Event

arXiv: DeepSeek@Jiaqi Su et al. (8)2 sourcesSource

Papers Builders09:28

Chinese contentDesign Anything：用意图-执行图让 Agent 长流程设计不跑偏

做 Agent 长任务的朋友看看这篇：把设计意图存图里、用确定性控制器核对交付物，19 个领域、228 个任务的基准测下来，同一模型约束满足率高出基线 14-25%。

#Design Anything #DeepSeek #GPT #智能体

arXiv: DeepSeek@Xuchao Ren et al. (11)Source

Papers Builders09:27

Chinese contentMetaSkill：用 DAG 让 Agent Skills 的组合编排可执行

Skills 之间怎么传递输出一直靠模型自己猜，这篇把依赖关系做成 DAG 直接可执行，做 Agent 编排的可以看看实现思路。

#MetaSkill #Agent #DAG #GLM-5.2

arXiv: DeepSeek@Meng Shu et al. (6)Source

Papers Builders12:37

Chinese content网易 LobsterAI 发布《中国办公 Agent 用户行为报告》：微信成最大入口

网易 LobsterAI 拿真实用户数据写的报告：微信是最大入口，七成人只用一个模型，跟 X 上的玩法差别挺大，做 Agent 产品的可以看看。

#LobsterAI #DeepSeek #GLM #智能体

Event

向阳乔木@vista82 sourcesSource

Papers Builders10:28

Chinese contentPresenceBench：评测大模型对用户状态的适应能力

四个主流模型谁更懂“你现在的状态”？这篇论文用五臂对照实测了，结果比想象中微妙，做对话产品的可以看看评测方法。

#PresenceBench #Claude #DeepSeek #Gemini

arXiv: DeepSeek@Peña Arroyo Gallego et al. (5)Source

Papers US vs ChinaSelected16:33

DeepSeek V4.1 RL environment counts still hard to compare

RL environment counting has become a reporting gap: different labs may describe the same training run with wildly different sandbox numbers, making cross-lab comparisons unreliable. Practitioners tracking frontier RL scaling should read both the DeepSeek paper and the Beam blogpost before citing either figure. The expectation that V4.1F-1017 outperforms Beam sets up a concrete head-to-head comparison to watch.

#DeepSeek #Reflection AI #RL environments #reasoning models

Teortaxes@teortaxesTexSource

Papers US vs ChinaSelected04:16

SWA Attention: DeepSeek and Kimi Chose Independent Architecture R&D Instead

The takeaway is that attention architecture choice has become a strategic fork: American spinouts standardise on SWA while Chinese labs treat attention design as proprietary IP. Engineers evaluating open-weight models should look past parameter counts and check which attention variant a model uses, since it shapes cost and context behaviour. The comparison with Mistral's stack gives a concrete reference point for judging the trade-offs.

#DeepSeek #Kimi #Mistral #open-source models

Teortaxes@teortaxesTexSource

Papers09:38

Chinese content联邦微调中的隐私泄露与后门投毒研究：FedAvg 攻击与 FLTrust 防御实测

arXiv 上的联邦学习攻防实验报告，服务端能从一次更新里偷出标签集，30% 恶意客户端就能植入后门，FLTrust 能防住，代码可直接跑。

#FedAvg #FLTrust #联邦学习 #数据投毒

arXiv cs.LG@Vasileios Perifanis, Nikolaos PavlidisSource

Papers BuildersSelected09:28

IneqAgent Pairs LLM Agents with Symbolic Backends for Lean4 Inequality Proving

The interesting part is the division of labour: the LLM handles high-level lemma strategy while exact symbolic backends supply certificates, keeping every step Lean-verified and cheap at low inference budgets. Formal-methods researchers and theorem-proving engineers can reuse the MCTS-plus-backend pattern for other proof domains. It also gives DeepSeek-V4-Flash a concrete reasoning benchmark data point against other backbones tested in the paper.

#IneqAgent #DeepSeek-V4-Flash #Lean4 #reasoning models

arXiv: DeepSeek@Gaolei He et al. (5)Source

Papers BuildersMulti-source09:17

Chinese content研究测试向 Qwen3.8-Flash 与 DeepSeek-V4.1 记忆模块写入知识的可行性

把知识写进模型记忆到底行不行？这篇论文在 Qwen 和 DeepSeek 上做了实测，命中率数据挺扎眼，做记忆机制研究的值得看数据。

#Qwen3.8-Flash #DeepSeek-V4.1-Flash #GPT-2 #知识注入

Event

arXiv: DeepSeek@Wei Gao et al. (4)8 sourcesSource

Papers BuildersMulti-source13:00

Chinese content研究拆解 LLM 网络威胁情报实体抽取失败模式：主要问题不是幻觉

DeepSeek-V4-Flash 和 GPT-6.1-Sol 抽实体都不靠谱，但坏的原因不是幻觉，而是把背景信息当成攻击实体。用置信度分层排审核优先级，别当真值。

#DeepSeek-V4-Flash #GLM-5.3-Flash #GPT-6.1-Sol #网络威胁情报

Event

arXiv: DeepSeek@Jiahong Qian5 sourcesSource

Papers BuildersSelected12:52

BizBench Benchmark Tests Agent Execution Validity in Payments Workflows

The benchmark shows that outcome-only evaluation masks invalid execution, since models can hit the final goal while violating constraints mid-run — a gap that matters most in regulated domains like payments where intermediate errors carry financial or compliance consequences. Agent builders can use the 50-task suite and deterministic checks to measure constraint compliance and tool-call overhead, not just task success. Teams choosing models for workflow automation get a direct comparison of frontier and open models under identical conditions.

#BizBench #agents #benchmark #Claude Opus 5

arXiv: DeepSeek@Enhan Zhao et al. (9)Source

Papers Builders12:48

Chinese contentAtomSearch：用证据驱动的树搜索优化 Agent 技能

一篇讲怎么高效改 Agent 技能规则的论文，AtomSearch 在 deepseek-v4-flash 上比 SkillOpt 高 9.68 个百分点，技能还短好几倍。

#AtomSearch #deepseek-v4-flash #SkillOpt #智能体

arXiv: DeepSeek@Jia Tong et al. (4)Source

Papers Builders12:46

Chinese content论文提出 Memory Echo 现象与 FPA 框架，修复智能体长期记忆失真

做智能体长期记忆的可以看看：论文发现同源旧记录会重复刷存在感，FPA 框架实测把准确率拉高近 10 个点，代码开源可复现。

#Memory Echo #MEchoBench #DeepSeek V4 Flash #Kimi K2.6

arXiv: DeepSeek@Yufeng Li et al. (8)Source

Papers Builders12:44

Chinese contentarXiv 论文：证伪驱动的模型无关智能体协议，输出纪律决定决策下限

用中医辨证基准证明了件事：很多模型翻车不是能力不够，是输出纪律不行。他们设计的证伪协议能把 Kimi K3 的方差从 0.218 压到 0.007，写智能体的值得看看。

#TCMEval-SDT #DeepSeek-V4-Flash #Kimi K3 #智能体

Event

arXiv: DeepSeek@Jing Liu et al. (4)2 sourcesSource

Papers Builders12:41

Chinese contentANIS 框架为智能体生成原生桌面界面，成功率达 74.5%

给智能体设计界面这事有论文了，DeskIF-Bench 和 ANIS 都开源思路，做 agent 桌面自动化的可以看看怎么提分还省 73% token。

#DeskIF-Bench #ANIS #DeepSeek #Claude Haiku 4.5

Event

arXiv: DeepSeek@Alan Wu et al. (3)2 sourcesSource

Papers Builders12:37

Chinese contentSeTPO 训练方法让小语言模型更擅长调用外部工具

小模型学用工具的老难题有了新解法：SeTPO 用双向 KL 同时管住探索和模仿，BFCL v4 上跑赢了蒸馏和纯 RL，做端侧模型的可以看看。

#SeTPO #BFCL #API Bank #小语言模型

arXiv cs.LG@Haoxuan Gao et al. (8)Source

Papers Builders00:53

Chinese contentSRO：用LLM智能体自动优化后训练配方，Qwen3上远超基线

SRO让LLM自动设计后训练配方空间，再用贝叶斯选试验，Qwen3上分数远超基线，代码已开源。

#SRO #Qwen3 #智能体 #后训练

arXiv: DeepSeek@Ruohong Liu et al. (5)Source

Papers Builders11:35

Chinese contentBRQGen：迭代式评估 Bug 报告完整性并按需生成追问

这篇提出了一个挺聪明的思路：不再机械地问固定数量的问题，而是按报告缺口按需追问，作者还放了 429 条真实数据集和多个模型的对比成绩。

#BRQGen #DeepSeek V4.1 #Qwen #软件工程

arXiv: DeepSeek@Tevfik Emre Sungur et al. (4)Source

Papers Builders17:55

Chinese content研究：SWE-bench 上通过测试的补丁近半数存在运行性能差异

四个前沿模型解 SWE-bench，通过测试的补丁一半有性能差距，模型性能排名还和解决率反着来，做工程的人可以看看。

#SWE-bench #SWE-agent #Claude Opus 5 #GPT 5.6

arXiv: DeepSeek@Han Peng et al. (5)Source

Papers Builders09:52

Chinese contentSCCGBench：真实Web服务调用代码生成评估基准

SCCGBench评估了六种代码LLM在真实Web服务调用中的表现，揭示了API接地是主要瓶颈。

#SCCGBench #RapidAPI #GitHub #代码生成

Event

arXiv: DeepSeek@Shuo Wang et al. (5)2 sourcesSource

Papers BuildersSelected14:42

Internalizer maps context to parameters for DeepSeek V4-Flash

The shared-expert placement means one hypernetwork can steer every layer's routed experts at once, which sidesteps the usual per-layer adapter sprawl and could make context injection cheaper at inference time. Teams running Mixture-of-Experts models such as DeepSeek V4-Flash or Qwen variants should watch whether the mapping survives fine-tuning and long-context loads before adopting it over LoRA.

#DeepSeek V4-Flash #hypernetwork #fine-tuning #open-source models

Teortaxes@teortaxesTexSource

Papers US vs ChinaSelected04:29

Tsinghua, Oxford, Stanford Paper Finds LLMs Reason Even When Told Not To

The finding that a missing think tag or short reply does not mean a model skipped reasoning undermines a common evaluation shortcut. Teams benchmarking reasoning models or building answer-only constraints should re-check how compliance is measured, since suppressing visible reasoning costs roughly 15 accuracy points. The result spans 5 models, making it a broad behavioral pattern rather than a single-vendor quirk.

#DeepSeek #DeepSeek-V4-Flash #reasoning models #AI safety

Event

rohanpaul_ai@rohanpaul_ai2 sourcesSource

Papers BuildersMulti-source10:23

Chinese contentAgenticBBO-Bench：评测 LLM 智能体黑盒优化能力的跨域基准

想用 LLM 智能体做黑盒优化的可以看看，五个领域统一评测，还拆了哪些设计选择真有用、哪些是伪收益。

#AgenticBBO-Bench #LLM #GPT-6 Astra #DeepSeek-V4.1-Flash

Event

arXiv: DeepSeek@Ming Chen et al. (11)11 sourcesSource

Papers Builders10:23

Chinese contentInternalizer 超网络可将文档编译为 284B 参数模型的 LoRA 适配器

把一份文档直接编译成 LoRA 权重，284B 的 DeepSeek v4 Flash 也能用，top-1 准确率从 63.4% 拉到 84.9%

#Internalizer #DeepSeek #LoRA #超网络

arXiv: DeepSeek@Peter Devine et al. (4)Source

Papers BuildersMulti-source10:22

Chinese contentMine Odyssey 基准用 Minecraft 重建真实地点评测空间智能体

拿 Minecraft 重建曼哈顿、白金汉宫这些真实地点来考模型找路，GPT-6 Astra 考了 85.6 分，开源的 DeepSeek-V4.1-Flash 才 23.9，差距很直观。

#Mine Odyssey #GPT-6 Astra #DeepSeek-V4.1-Flash #智能体

Event

arXiv: DeepSeek@Yuxuan Cao et al. (4)11 sourcesSource

Papers BuildersMulti-source10:21

Chinese content评估LLM生成评分标准的干预转移方法

论文提出了一种创新的评分标准评估方法，揭示了不同LLM生成评分标准之间的不对称性，对AI性能监控有重要启示。

#评分标准 #干预转移 #LLM评估 #HealthBench

Event

arXiv: DeepSeek@Erik Skalnes et al. (4)5 sourcesSource

Papers BuildersMulti-sourceSelected08:54

ByteDance Seed finds phase sensitivity behind DeepSeek long-context drift

Benchmark averages can mask periodic failure modes in long-context retrieval, so a model that looks solid on mean scores may still drop retrievals depending on where content sits in the compression window. Engineers running DeepSeek or similar KV-compressed models on RAG or agent pipelines should test retrieval across phase positions rather than relying on aggregate accuracy. The arXiv paper (2609.36322) offers the evaluation methodology to replicate.

#DeepSeek #ByteDance #KV cache #long-context

Event

IT之家6 sourcesSource