Jiqizhixin (机器之心)
AI news source. Articles are auto-selected and adapted by Hamidun News editors.
Latest publications
CAS Institute of Software and Kuaishou Build First Benchmark for LLM User Simulation
The CAS Institute of Software and Kuaishou introduced a NeurIPS 2026 benchmark testing whether LLMs can convincingly play a real user across the entire behavior chain.
RSIGame: a 27B model combo plays its own games and outperforms GPT-5.5
China's RSIGame has AI test, fix and refine games on its own. The authors claim a combination of 27B models beat GPT-5.5 in one-shot generation.
Six Teams from UCSB, Stanford and J.P. Morgan Release an Agent Memory Benchmark
WorldMemArena tests how multimodal agents remember and use what they have seen in real interactions. The six-organization paper has been accepted to NeurIPS 2026.
九章云极 Alaya-DFlash2 speeds up GLM-5.3 decoding by 3.41x
Chinese cloud provider 九章云极 unveiled Alaya-DFlash2: speculative decoding speeds up GLM-5.3 by 3.41x on long inputs, such as 32k-token documents.
MemT 1.0 builds memory into the model and aims at AI self-improvement
MemT 1.0 is out: memory is meant to be a built-in model capability, not an external add-on. A story about DeepSeek and an exhausted context sparked the discussion.
Huawei openJiuwen connects an LLM to lab equipment for electrocatalysis
Huawei's openJiuwen team and its computing platform team closed the loop: an LLM designs an electrocatalysis experiment, and physical instruments carry it out.
Anhui Bets on Humanoids: A Robotics Park in Hefei and a 15.1 Billion Yuan Project
Hefei has opened a robotics park with 14 companies, and Anhui is now citing 31 humanoid models and a 15.1 billion yuan project. Here is what we know about China's robotics cluster.
Jiqizhixin Invites NeurIPS 2026 Authors to an Offline Meetup in Beijing
Chinese outlet Jiqizhixin has announced an in-person event in Beijing where readers can question NeurIPS 2026 authors directly about world models, embodied AI and RSI.
Xiaohongshu speeds up its sparse-attention indexer: DoSA halves TTFT
Xiaohongshu's Dots team proposes DoSA, which prunes about 97% of candidates in the indexer, keeps the same Top-K and cuts time to first token by half.
AstraZeneca to pay insitro up to $1.5 billion for AI-driven MASH drug discovery
AstraZeneca and insitro signed a deal worth up to $1.5 billion: the Virtual Human platform will search for drugs for MASH and cardiomyopathies. It is one of the largest deals in AI pharma.
Dream Labs, from the NVIDIA DreamZero team, raises $90M and opens an API for robots
The DreamZero team, formerly at NVIDIA, has emerged from stealth with a $90M seed round and a public Dream API that controls a robot from a text command.
UniPat AI to Put 100 Million Yuan Into Fellows Program for Scientists
Chinese startup UniPat AI is launching UniPat Fellows: 100 million yuan over one year, 500,000 yuan in API credits per researcher, plus data, experts and compute.
USTC Releases "Machine Scientist 2.2": AI Moves From Running Experiments to Creating Substances
USTC has unveiled its Machine Scientist 2.2 series. The claim: the system doesn't just run experiments but autonomously creates new materials.
First Global Embodied AI Challenge Closes Alongside China Robot Competition 2026
The first global embodied AI application challenge and China Robot Competition 2026 wrapped up in China on October 18. A CAS academician and Binjiang officials attended the closing.
Four Chinese Players Are Building Real-Time Interactive Video Worlds
Alibaba's HappyOyster, Ant's LingBot, Vidu S2 and PixVerse R2 put interactive video worlds at the center of China's AI market: the model reacts to user actions as it generates.
Ant Group releases Ling-3.1-flash: agent runs for 10+ hours straight
Ant Group unveiled Ling-3.1-flash from its Bailing line. According to Jiqizhixin, the model runs agentic tasks for more than ten hours and can handle building a city.
Hefei's Zhongke Ion Raises 600 Million Yuan for Proton Cancer Therapy
A Hefei company whose accelerator grew out of "artificial sun" fusion technology has raised 600 million yuan. Its domestically built system is already treating patients.
LabBuilder Turns an Experiment Description into a Safe 3D Lab
A team at ICML 2026 introduced LabBuilder, a system that builds a safety-aware 3D lab layout from the text of an experiment protocol. The authors name three gaps in existing methods.
Hesai CEO: A Robot in a DQ Ice Cream Shop Needs More Than Hands and a Model — It Needs to Understand Space
At a Hesai presentation, CEO Li Yifan explained why a robot in a DQ ice cream shop is limited not by its manipulator but by space. The headline promises scene reconstruction in 15 minutes with one camera.
Codex Climbed Out of Its Slump Under Tibo: Every Chronicles a Year-Long Comeback
Every founder Dan Shipper spent months interviewing sources for his report "How Codex Happened." Chinese media retell it as the story of how Codex "swallowed" ChatGPT.
Two Simple Changes Nearly Double the Correlation of MLLM Encoder Scoring
Chinese researchers propose two simple improvements to pre-evaluating vision encoders for multimodal LLMs. Correlation with final quality nearly doubled.
STILL from HIT Shenzhen: Hybrid Attention Within a Layer Picks Tokens by Content
Researchers at Harbin Institute of Technology, Shenzhen, present STILL at NeurIPS 2026: a content-dependent hybrid of softmax and linear attention inside a single layer.
Robots learn from 20 854 hours of video: how Sharpa is building dexterous manipulation
Robotics data is moving from costly teleoperation to human video: EgoScale used 20 854 hours. Here is how Sharpa fits its system into that logic.
Xiamen University's FlowMAS Teaches Teams of AI Agents to Assign Their Own Roles
ASC Lab unveiled FlowMAS under the NeurIPS 2026 banner: a system meant to teach teams of LLM agents to organize their own work. The source has no metrics yet.
Hefei Startup Raises Tens of Millions to Turn Mega-Facilities into 3D Data for AI
A Hefei company raised tens of millions to convert large scientific facilities like the EAST tokamak into 3D data that AI models can read.
UBTech's Thinker-WM Takes Three First Places, Scoring 0.740 on WorldOlympiad
UBTech's Thinker-WM world model topped three embodied AI benchmarks. It scored 0.740 on WorldOlympiad, and a single model covers six capabilities.
Google DeepMind Launches AI Agent to Hunt for Antimicrobial Peptides Among Millions of Proteins
DeepMind has shown an agent that sifts through millions of protein sequences and selects antimicrobial peptide candidates for experimental testing.
Synchronous MoE decoding makes GPUs wait for the busiest group
In MoE deployments like DeepSeek, every decoding step runs in lockstep, so the most heavily loaded GPU group holds back the rest. Throughput grows more slowly than the number of cards.
A Company from Nanjing University's Orbit Bets on a Fourth Driver of AI Growth
A team from Nanjing University's orbit sees a fourth driver of AI growth after compute, algorithms and data: connectivity and cooperation among independent models.
Tadpole Proposes a Foundation Model for Physics: From Prediction to Understanding
Tadpole brings foundation-model logic to science: one model instead of separate networks for weather, airflow, material deformation and electromagnetism.