Tsinghua University
AI overview
Sign in and the AI will write an overview from our coverage.
Headlines · 23
- Tsinghua prize winner's world-model startup reported at $600M valuation after three rounds
Wu Haixu, a 2022 winner of Tsinghua University's top scholarship, has founded a world-model startup called Wuyuan Technology (物原科技), which has already closed three funding rounds and reached a reported valuation of about $600 million (roughly RMB 4 billion), according to QbitAI. Investors named in the report include Ant Group, Meituan, Dachen Caizhi and Xinglian. The company is said to focus on world models with physical modeling capability, partly extending GeoPT — a February paper Wu co-authored as first author that uses synthetic dynamics to lift static geometry into dynamic space so models learn physical intuition from particle trajectories without labels. Wu is currently a postdoc at MIT CSAIL with Wojciech Matusik; GeoPT co-authors also included Tsinghua associate professor Long Mingsheng and MIT's He Kaiming. The valuation and deal details come from people familiar with the matter and have not been confirmed by the company.
36氪 人工智能 · 🔥 13 - AgiBot to invest over ¥200M with 20+ universities on embodied AI talent
AgiBot said its "YuanSheng ecosystem development plan" has signed its first batch of more than 20 universities—including Peking University, Tsinghua, Zhejiang, South China University of Technology, Shanghai Jiao Tong, Xi'an Jiaotong, Southeast and Sun Yat-sen—to build joint embodied-AI teaching and research platforms. The company will supply more than 1,000 embodied-AI devices spanning its Genie, Lingxi and Expedition wheeled, half-size and full-size humanoids plus KuoTu quadrupeds, and provide its AIMA development platform with LinkSoul, LinkCraft and Genie Studio for curriculum, research and hands-on training; it values the hardware and software memberships at ¥200 million. AgiBot had earlier said the plan would spend over ¥2 billion over five years, and says it has connected with 237 research institutions, 500-plus ecosystem partners and more than 10,000 application developers.
IT之家 AI · 🔥 5 - Tsinghua open-sources Puro-2B: 2B pretraining on RTX 5090s for about $4,400
A Tsinghua University team led by professors Chen Wenguang and Lyu Kaifeng released Puro-2B, a fully open pretraining recipe that publishes the training data, framework, configuration and model weights. The roughly 2B-parameter model was trained from scratch on a cluster of consumer RTX 5090 GPUs using up to 1.4 trillion tokens; on a 15-task evaluation, a version with about $4,400 of pretraining compute beat Qwen2-1.5B on average, while the $6,900 best version came close to Qwen2.5-1.5B on average. The team also fit a Puro Cost Scaling Law to estimate achievable quality at a given budget, combining hardware choice, FP8 training, the MuonH optimizer, Curriculum Model Averaging and data-slice evaluation.
36氪 人工智能 · 🔥 8 - Tsinghua, BUPT paper unveils DIVA jailbreak attack on diffusion VLMs
Tsinghua University and Beijing University of Posts and Telecommunications propose DIVA (Discrete-diffusion Vision-language model Attack), a white-box visual jailbreak method aimed at discrete diffusion vision-language models (dVLM), which re-use visual conditioning at every denoising step — an effect the authors call Cross-Step Conditional Propagation. The method splits a harmful intent across a composite image and a rewritten prompt, then jointly optimizes a bounded pixel perturbation over multiple denoising timesteps as a category-level, offline shared perturbation. On the HADES benchmark (Beaver semantic metric), average attack success rates were 58.80% for LLaDA-V, 67.73% for MMaDA and 69.07% for LLaDA2.0-Uni; on VLSBench they reached 52.07% for LLaDA-V and 74.16% for MMaDA.
新智元 · 🔥 5 - Vidu Q4 Preview debuts with stronger acting, camera work and 0.09 yuan/sec pricing
Shengshu Technology released Vidu Q4 Preview, a video generation model emphasizing character performance, dynamic camera movement and complex visual effects across 2D and 3D animation scenes. It produces clips up to 16 seconds, supports 2K and 4K output, and accepts up to 15 reference images and 3 reference audio tracks. The company says API image-to-video costs about 0.6 yuan per second at 720p and 0.75 yuan at 1080p, with a two-month launch promotion on the Vidu platform from 0.09 yuan per second, and that a 5-second clip can be generated in as little as roughly 110 seconds. The report also notes Seedance 2.5's push into 30-second long-form narrative and editing control, and Kling's continued focus on realism and fine-grained control.
新智元 · 🔥 3 - Smart electric yacht startup MIRAVA raises nearly RMB 50M angel round
MIRAVA, a Chinese maker of smart electric consumer boats, has raised close to RMB 50 million in an angel round led by Jiuhe Venture Capital, with Chuxin Capital and Eversnow participating; the funds are earmarked for core technology R&D, mass-production line construction, first customer deliveries and a localized European channel and service network. According to the company, its first 8-meter smart electric yacht, the M800E, was unveiled for Europe at the Cannes Yachting Festival in September 2026, is slated for boot Düsseldorf in January 2027, and the first founder-edition units are due for delivery in mid-next year, with production already moving into small-batch delivery. MIRAVA says the M800E uses a smart wristband as a positioning beacon for scenario-based autonomous follow-me operation and adds a detachable rear sub-boat module, positioning the product for casual "water players" rather than traditional boating enthusiasts.
36氪 人工智能 · 🔥 3 - Venture panel: Physical AI hits data, safety and long-horizon hurdles
At a panel during the 5th Global Digital Trade Expo's "Digital Trade Venture Day" (Sept 22-24, Hangzhou), moderator Liu Huining of Huiyan Future and Tsinghua University, Puhua Group chairman Cao Guoxiong, Qiming Venture Partners investment partner Shen Jin and Gaoshan Xinyu founding partner Wang Xiang debated how Physical AI reaches industry. Cao said the field is still in a validation phase: demos work, but closed loops in real scenarios are unproven, with hotel delivery already solved while cleaning and hazardous-environment inspection are still in progress, and a humanoid form factor not necessarily required. Shen listed six infrastructure pillars — compute, data engineering, models, simulation and synthetic data, operating environment including safety, and body hardware — and said safety may be the final obstacle; Wang estimated physical-world data needs could be orders of magnitude larger than the hundreds of billions of tokens used by large language models, making real-world data collection critical. Cao added that a closed loop for an embodied "brain" may take at least five years.
创业邦 科技 · 🔥 0 - PAIR 2026 roundtable: who verifies AI-generated scientific findings?
PAIR Conference 2026, themed on the frontier of next-generation physical intelligence, was held in Shanghai on Sept 20–21, and its "next-generation scientific discovery" forum closed with a roundtable on how model outputs become trustworthy scientific evidence — cross-disciplinary verification, reproduction and applicability limits. Panelists said AI generates candidates far faster than experiments can check them: the best AI virtual-cell data covers only about 100 perturbed genes out of roughly 20,000 human genes, GNoME's ~2.2 million predicted crystal structures were criticized as largely simple extensions of known compounds, and A-Lab corrected its "new materials" wording after re-analysis of its phase identification. Their proposed principles: fix verification standards before experiments and independently of results, separate the generator from the verifier, use independent labs and cross-platform replication, and require results to be independent, prospective and extrapolative.
MIT科技评论中文 · 🔥 0 - Nature Will releases Natura-1: 50B MoE model for dexterous robot hands
Beijing embodied-AI company Nature Will released Natura-1, a 50B-parameter MoE Transformer physical-world foundation model pretrained on roughly 1 million hours of first-person human experience data (about 61.7 million clips, 318,000 tasks, 14,000 scenes), then post-trained on limited robot data; it takes language and vision as input and drives 20-degree-of-freedom dexterous hands within a 100 ms control loop. The team says it combines dexterous manipulation with zero-shot instruction following, and published unedited one-take demos including wrapping wontons, making tea, packing parcels, applying a phone screen protector and writing calligraphy. On its own 12-dimension benchmark, Natura-1 scored 53.4 (Base) and 59.1 (Agentic) on the dexterous-hand platform, while π0.5 and OpenAI's GPT-6 Astra scored zero there (13.5 and 16.7 on a gripper platform); the team also reports action-prediction error dropping 47.7% on held-out human data and 67.7% on held-out robot data as training experience scaled from about 44,000 to 1 million hours.
机器之心 · 🔥 0 - VPP2 and Dyna-2: how human video trains robots
A TMTPost feature by Jushenzhi reports that on October 10 Robot Era open-sourced two sets of VPP2 video model weights; VPP2 splits manipulation learning into predicting how a task unfolds, compressing it into visual plans and learning robot actions, and only the first two stages were released. Dyna Robotics' Dyna-2 is described as a scale test: four models pretrained on 1,000, 10,000, 100,000 and 1 million hours of first-person human video then received identical robot-task training, the first three failed to turn a key while the 1-million-hour model succeeded 9 of 10 times. Nvidia's DreamZero added 12 minutes of action-label-free human video for nine previously unseen tasks and lifted average task progress from 38.3% to 54.3%, while RoboWM-Bench from Peking University and others converted generated video into robot actions in simulation, reporting 50%, 20% and 0% success for Wan 2.6 on closing drawers, picking objects and placing objects. The article argues human video is taking over task understanding and world dynamics while scarce robot data focuses on body control.
钛媒体 · 🔥 0 - PAIR 2026 panel maps how AI and quantum computing enable each other
PAIR Conference 2026, themed "Frontiers of Next-Generation Physical Intelligence," was held in Shanghai on September 20–21, and its next-generation computing paradigms roundtable discussed two-way empowerment between quantum computing and AI. Hongye Hu of the Institute of Physics, Chinese Academy of Sciences, said Chinese teams have achieved automated calibration of 100-plus-qubit chips without continuous human involvement, and cited Google's AI-controlled physical device achieving non-stop quantum operations, while noting reinforcement learning needed 10^5–10^6 samples versus far fewer with physics-informed methods. Anbang Wu of Shanghai Jiao Tong University backed AI for quantum compiling, program verification and error-correction optimization; Zheng Yan of Westlake University said publication growth in strongly correlated quantum many-body work has not been similarly lifted by AI. On the reverse direction, Hu described quantum generative models with a proven unconditional polynomial memory advantage in certain tasks, while Wu said he has not yet seen a new AI paradigm that truly exploits quantum advantages.
MIT科技评论中文 · 🔥 0 - Carnegie report: China tops US as the No. 1 workplace for elite AI researchers
A Carnegie China report released on 23 September, "Who's Ahead in the Global AI Talent Race?", tracked the CVs of 10,280 authors of NeurIPS 2025 papers and found that about 41% of top AI researchers worked in China in 2025 (the report's figure is 40.6%) versus 34% in the US (34.2%), compared with 27% against 46% three years earlier. The same report says the US still had a net inflow of 2,145 researchers while China had a net outflow of 1,729, that China's retention of its own top researchers rose from 57% to 69%, and that roughly 98% of China-based top researchers did their undergraduate studies in China versus only about a third of US-based ones in the US. Peking University topped the institutional ranking for the first time, followed by Tsinghua University, KAIST and Shanghai Jiao Tong University, while 2022 leader Google slipped amid a reported six-month publication ban on strategic generative AI papers; the report acknowledges its sample skews academic and may undercount industry.
虎嗅 AI · 🔥 0 - AI PhD graduates draw packages up to 8M yuan as Chinese tech giants battle for talent
A widely circulated social-media post claims NVIDIA offered a 2026 Peking University integrated-circuit PhD a first-year pre-tax package of about 2.9 million yuan, including signing and performance bonuses plus a possible Silicon Valley posting after two years; the report notes that NVIDIA China PhD campus offers were generally 450,000–600,000 yuan five years ago. Recruiters cited in the piece say multi-million-yuan packages for top-school AI PhDs are no longer unusual, with Tsinghua's Institute for Interdisciplinary Information Sciences large-model graduates starting at 4–5 million yuan, JD.com offering 8 million yuan to a new graduate and Xiaohongshu about 6 million yuan to a master's graduate in large models. ByteDance, Tencent, Alibaba and Kuaishou run programs such as Top Seed, Qingyun Plan, Alibaba Star and Kuaishou Star, with ByteDance's 2026 Top Seed cohort recruiting roughly 30 PhDs; the article also cites market chatter of a near-160 billion yuan ByteDance AI budget for 2026 and Alibaba's 380 billion yuan three-year AI and cloud commitment, plus over 100 billion yuan of quarterly Tencent AI infrastructure spending.
36氪 人工智能 · 🔥 0 - Infinigence AI raises 2.3B yuan Series C, total funding reaches 4.3B yuan
Chinese AI infrastructure company Infinigence AI has closed a 2.3 billion yuan Series C round, bringing its cumulative funding to 4.3 billion yuan — a record for both a single round and total funding in China's AI infrastructure sector, according to multiple outlets. Founded in 2023 and incubated by the Tsinghua University electronics department's NICS-EFC lab, the company avoids building chips or training foundation models, instead working on heterogeneous compute scheduling and hardware-software co-optimization, and says its platform reaches 37,000P of compute across 16 mainstream chip types. It says daily token calls on its Agentic MaaS platform grew 40-fold from December, and that its cross-cluster heterogeneous PDD architecture cut first-token latency by 51.5% while reducing per-token cost by 37.5%; Bloomberg, citing people familiar with the matter, reported the company filed confidentially for a Hong Kong IPO.
智东西 · 🔥 0 - Tsinghua-linked Xianyin Weilai raises hundreds of millions for rope-driven space robots
Anhui-based Xianyin Weilai Technology, which says it was founded in September 2025, disclosed that it closed three funding rounds — angel, Pre-A and Series A — within four months between May and September 2026, raising several hundred million yuan in total, according to a Hard Kr report. Lanchi Ventures, Hillhouse Ventures and Binfu Capital led the angel round, Meridian Capital China led the Pre-A round, and Alibaba together with a China Mobile-affiliated fund led the Series A, with industrial investors including AgiBot and Trina Solar also taking part. The company says its rope-driven flexible space robots target in-orbit refueling, assisted de-orbiting and in-orbit inspection and maintenance, and it plans a space rendezvous-and-docking technology demonstration in 2027.
36氪 人工智能 · 🔥 0 - Yau leads math-and-AI lab alliance at Tsinghua with 1,000 researchers
On October 11 the Yau Mathematical Sciences and AI Laboratory Alliance was launched at Tsinghua University, led by Fields Medalist Shing-Tung Yau and formed by eight institutions including Tsinghua's Qiuzhen College, the Yau Mathematical Sciences Center, the Beijing Yanqi Lake Institute of Applied Mathematics, the Morningside Center of Mathematics and the Chinese University of Hong Kong's Zhen-Cross Mathematical Sciences Academy, aiming to bring together roughly 1,000 mathematicians and researchers. The alliance will work on math-specific large models, AI-assisted formal proofs and conjecture discovery, and interpretable AI, with labs planned in Beijing, Shanghai, Nanjing, Shenzhen and Hong Kong that share model-training resources, formal math libraries and compute platforms, plus a planned AI-assisted mathematics journal. Yau described AI's impact as a 'tsunami' that researchers should ride rather than avoid, and said China's own large models are still not world-leading in frontier mathematics and basic science.
虎嗅 AI · 🔥 0 - AI makes knowledge cheap: what universities still sell
In a paper posted to SSRN titled "The University after AI," Alfaisal University economics professor Jason Potts argues that universities never really sold knowledge — they sell matching and verification, and AI is the first information technology to hit both at once. Data cited in the piece puts the US college wage premium at about 87% in 2022 falling to about 78% in 2026, while a Harvard Business School and Burning Glass Institute study of 11,300 roles found that dropping degree requirements raised actual hiring of non-degree holders by only about 3.5%, with roughly 97,000 Americans gaining jobs that way in 2023. The article also covers responses such as revived oral exams and separating teaching from certification, and notes that China's publicly funded universities have more of a buffer even as a projected 12.7 million graduates in 2026 face a further diluted degree signal.
MIT科技评论中文 · 🔥 0 - Ma Huateng sits beside Tencent chief AI scientist Yao Shunyu at Shenzhen briefing
Chinese media report that on October 8 Guangdong party secretary Huang Kunming visited Tencent's Shenzhen headquarters to review its AI research and industrial plans, and that chairman and CEO Ma Huateng sat next to Tencent chief AI scientist Yao Shunyu at the following discussion. The reports say Yao, born in 1998, studied in Tsinghua University's Yao class and earned a computer science PhD at Princeton before joining OpenAI as a research scientist in 2024, where he worked on the operator agent product and deep research and is linked to the ReAct method. He left OpenAI for Tencent in September 2025 and in December was named chief AI scientist in the CEO/president's office, reporting to president Liu Chiping while also heading the AI Infra and large language model departments. TMTPost says Tencent merged its Hunyuan multimodal and LLM departments into a foundation model department under Yao in July, and cites the half-year report for capital expenditure up 81.9% year on year to 84.72 billion yuan.
DoNews · 🔥 0 - Nikon strips Small World in Motion winner of title over generative AI use
Nikon disqualified Dr. Ning Xu of Tsinghua University from first place in its Small World in Motion microscopy competition, saying the video "did not comply with the competition rules regarding generative AI." The entry had claimed to show tiny hair-like cilia beating in the airway of a child with the respiratory condition PCD, but researchers and two former judges raised concerns after spotting structural abnormalities, cell features appearing and disappearing, and what one PhD student described as an AI watermark in the source material. Xu acknowledged on LinkedIn that an unsupervised neural-network method was used for "AI-assisted post-processing" of reconstructed grayscale images, while denying AI generated the movie, the cilia or their motion. Nikon, which said the decision is not a judgment of Xu's professional reputation or scientific contributions, replaced him with Vietnam's Nguyen Nam Nhat and said it will revisit the contest's rules and evaluation procedures.
BBC Technology · 🔥 0 - Embodied AI founders: tactile data, human-in-the-loop beat waiting for emergence
At a panel at the 2026 (20th) DEMO CHINA event, Wang Yu of Daimon Robotics, Mo Yilin of Lingyu Intelligence and Qin Shentao of OriginFlow discussed how embodied AI moves from research to real settings. Mo argued there is no need to wait for a large-model-style "emergence" moment: his company uses an L2-like human-in-the-loop mode, with robots already doing supermarket sampling, crab-catching and drink-mixing tasks, which he says can hit 70-80% of human efficiency on 70-80% of the work with human backup. Wang said force and tactile data cannot be replaced by video, while Qin collects human motion signals via non-invasive surface EMG and argued that data volume alone is not value — systems engineering and economic efficiency are the real moat.
创业邦 科技 · 🔥 0 - Global AI Chip Summit 2026: from TOPS to tokens as agents reshape silicon
The 2026 Global AI Chip Summit in Shanghai brought together more than 60 speakers across an opening session and three forums on large-model AI chips, agent inference chips and embodied-intelligence chips, focused on the computing chain from TOPS to tokens — covering architecture innovation, 3D stacking and near-memory compute, KV cache, new memory types, compute-in-memory, analog computing, RISC-V and token-factory operations. Several vendors used the stage to preview or detail products: Houmo AI said its M50 chip has reached 150+ design wins in half a year of mass production and that its next flagship A20 has taped out; Intel previewed a high-density cabinet with 128 or 192 Granite Rapids processors that it says can run 50,000 concurrent agents; Maitech said its 3D TPU has taped out and is expected back in November. Those figures are vendor claims and have not been independently verified.
品玩 实时要闻 · 🔥 0 - Robot Era's VPP2 tops RoboDojo embodied benchmark, beating GPT-6 Astra
Robot Era's self-developed world-action model VPP2 topped the RoboDojo simulation leaderboard with a 32.26% average success rate and 39.26 average score, ahead of the reported 22.48% and 28.97 for GPT-6 Astra. The team says it achieved this without extra data or augmentation such as Agent RSI, also ranking first on generalization, fine manipulation and memory, and has open-sourced the model. On real hardware, VPP2 ran 10 zero-shot tasks on an ALOHA dual-arm robot with a 58.5% average success rate versus 40% for π0.5, taking the best result in 9 of 10 tasks.
36氪 人工智能 · 🔥 0 - Zhengxing Innovation launches 24/7 Physical AI retail solution with H1 and C1 robots
At the 22nd Asia Pacific Retailers Convention and Exhibition (APRCE 2026), Zhengxing Innovation unveiled a Physical AI solution for open retail settings built around human-robot collaboration: the general-purpose bipedal humanoid H1, the wheeled-arm handling and shelf-restocking robot C1 designed for narrow aisles, and M1, a unified robot management platform handling task allocation, status monitoring and human takeover. The company says its in-house lightweight world-action model SLM-0.5 reaches 98.6% average task success on the LIBERO benchmark with 24x faster inference than mainstream approaches, and that its RLinf-based reinforcement-learning post-training system STEAM achieves 95% success on a designated-product restocking task. Commercial service is planned for 2027 via direct purchase or RaaS subscription, with joint testing alongside global retail chains.
量子位（原生 RSS） · 🔥 0
Experience and discussion from the community
Share my Tsinghua University experienceAsk about Tsinghua University
Nobody has shared their experience with Tsinghua University yet.