AI最尴尬的短板,中国科学院出手了
Merged summary
TL;DR — 中国科学院自动化所“知境”团队推出社会智能工程栈,通过 SoMBench、Zing 和 Actio,让 AI 的社会推理能力可评测、可训练,并能更可靠地用于真实交互。
- SoMBench 包含信念推断、情绪理解、社会规范和关系建模等 71 项任务;20 个参测模型的最高得分为 72.08%。
- Zing 针对模型弱项合成训练数据,结合两阶段微调、强化学习和 on-policy 蒸馏,以增强社会推理并减少遗忘。
- 团队报告称,Zing-27B 多模态版和 Zing-32B 文本版在五项基准上的平均分分别为 79.80 和 76.32,略高于其引用的 GPT-5.5 与 DeepSeek-V4-Pro。
- Actio 在推理阶段动态调度社会技能,并引入显式信念状态记忆、可复用推理经验及具有文化背景的检索能力。
注:其余来源涉及营销智能体、AI 行业动态、生成式广告、个人 AI 及 AI 制药,与该工作并非同一项目,未纳入合并。
Sources (8)
AI最尴尬的短板,中国科学院出手了
TL;DR - The Chinese Academy of Sciences’ Zhijing team unveiled an engineering stack for social intelligence in AI: the SoMBench benchmark, Zing models and training system, and Actio deployment architecture. It aims to make social reasoning measurable, trainable, and reliable in real-world interactions.
- SoMBench covers 71 tasks across belief inference, emotion understanding, social norms, and relationship modeling; the best of 20 tested models scored 72.08%.
- Zing uses weakness-targeted synthetic data, two-stage fine-tuning and reinforcement learning, plus on-policy distillation to reduce forgetting.
- Reported five-benchmark averages were 79.80 for multimodal Zing-27B and 76.32 for text-only Zing-32B, narrowly exceeding the cited GPT-5.5 and DeepSeek-V4-Pro results.
- Actio dynamically routes social skills, explicit belief-state memory, reusable reasoning experience, and culturally grounded retrieval during inference.
出海企业苦等的可信任AI营销产品,飞书深诺做出来了
TL;DR - Feishu Shennuo launched Marvy 2.0, a multi-agent system that coordinates the full overseas marketing workflow. It aims to make AI-driven decisions more trustworthy through traceable execution, human approval, and learning from campaign outcomes.
- AgentOS coordinates five agents for market research, media strategy, creative production, ad placement, and data analysis under a Supervisor Agent.
- Marvy combines general-purpose models with marketing data, specialized skills, linear programming, and reinforcement learning for tasks such as budget allocation.
- A Trace→Eval→Agent loop records workflows, evaluates execution and outcomes, and feeds lessons into future campaigns.
- Company-reported results include 74% higher optimization efficiency and 89% less human time in one Nigerian campaign; these claims were not independently verified in the article.
长鑫科技今日登陆科创板!员工:和我们没啥关系,更关心未来薪资、待遇;日方拆完宇树机器人,认输了;黄仁勋、马斯克就中国AI同日发声
TL;DR - This technology-news roundup highlights China’s growing AI and robotics position, including praise from Nvidia and Tesla leaders, Kimi K3’s launch, and Unitree G1’s engineering maturity.
- Jensen Huang said models such as DeepSeek and Kimi benefit the industry; Elon Musk predicted China could eventually lead AI.
- Kimi K3 reportedly has 2.8 trillion parameters and a 1-million-token context window, targeting long-horizon coding and knowledge work.
- Japanese engineers concluded after dismantling Unitree’s G1 humanoid that closing the technology gap quickly would be difficult.
- OpenAI and Anthropic reportedly sought tighter restrictions on open-weight models, while Microsoft, Nvidia, Meta, IBM, and others publicly backed open AI.
即梦联合抖音电商、艾菲奖发起 “AI 广告爆改大赛” ,共创经典广告新表达
TL;DR - 即梦AI联合抖音电商和艾菲奖发起AI广告创作大赛,鼓励用户用生成式视频工具改编经典品牌内容。该活动展示了AI视频在广告创意与品牌营销中的实际应用。
- 即梦App提供分身演绎、一键跟拍等功能,可生成12秒以内的短视频。
- Web端支持剧情反转、视觉特效和完整故事设定,面向12秒以上的视频创作。
- 大赛获得多个知名品牌授权,设置品牌、公益等五个创作方向。
- 作品将按创新能力、技术水平、故事设定及视听表达等维度评选。
切入千亿美元市场,破解百年镇痛难题,英矽智能提名AI发现的新型镇痛候选药物
TL;DR - Insilico Medicine nominated ISM9528, an AI-discovered, brain-penetrant, non-opioid oral pain drug, as a preclinical candidate. It targets a previously unlinked pain mechanism and showed promising efficacy, safety, and pharmacokinetics in animal studies.
- PandaOmics identified and prioritized the undisclosed “Target Z” using omics, literature, and multimodal data; knockout studies supported its role in pain.
- Chemistry42 generated and optimized ISM9528 for potency, selectivity, brain penetration, metabolic stability, and bioavailability.
- In rodent pain models, ISM9528 matched or outperformed pregabalin and showed longer-lasting analgesia than morphine in one intravenous study.
- The company reports wide safety margins in 28-day rat and dog studies and plans GLP toxicology in October 2026, followed by clinical trials in 2027.
智能体走向终端,个人AI时代正在到来
TL;DR - Qualcomm and IDC outlined a “personal AI” vision in which agents run across phones, PCs, glasses, and wearables using coordinated on-device, edge, and cloud computing. The shift could make AI more personalized and proactive while reshaping device architectures and ecosystems.
- Personal AI depends on persistent sensing, personal memory, privacy safeguards, natural interaction, and cross-device continuity.
- Agent workloads combine perception, memory, reasoning, and tool execution, requiring high-performance, low-power heterogeneous computing.
- Qualcomm highlighted its Oryon CPU, Adreno GPU, Hexagon NPU, sensor hub, memory, and unified software stack as an implementation foundation.
- IDC expects phones to serve as the central personal-AI hub, with glasses and other devices acting as additional contextual interfaces.
全球首个!上海AI制药新融资,英矽智能投了!
TL;DR - Byterna Therapeutics raised tens of millions of yuan in Pre-A funding to advance AI-designed circular mRNA therapies, including an in vivo CAR-T candidate and personalized cancer vaccines.
- Funds will support a clinical filing for a dual-target circular mRNA in vivo CAR-T therapy currently in an investigator-initiated trial.
- Byterna applies AI to neoantigen prediction and sequence design for personalized circular mRNA cancer vaccines.
- The company and investor Insilico Medicine plan to build an AI-native R&D engine spanning circular mRNA design through clinical translation.
- The platform combines AI design, circular mRNA, and targeted lipid nanoparticle delivery.
AI制药,国药集团大动作!
TL;DR - Sinopharm and Harbour BioMed formed a joint innovation consortium to develop drugs, combining AI-enabled antibody discovery with clinical development, manufacturing, and commercialization capabilities.
- The collaboration targets oncology and immuno-inflammatory diseases across the drug-development lifecycle.
- Harbour BioMed will contribute its Harbour Mice fully human antibody platform and AI-based antibody discovery capabilities.
- Sinopharm will support clinical trials, commercial-scale process development, clinical-sample manufacturing, and sales.
- Both companies will share R&D costs and product rights according to agreed proportions.