Hong Kong University of Science and Technology

Headlines · 5

1. 6d
   EdgeRock-4B Opens With Fixed-Memory Long-Context Inference

   Rock Matrix, Hong Kong University of Science and Technology, Carnegie Mellon University and other teams released and open-sourced EdgeRock-4B, according to QbitAI. The Qwen3.5-4B-based model uses recurrent state, a local window and Rock Memory Anchors to keep KV Cache at 72 MiB under a 256K context, reducing it by 99.1% from the source model. In team-reported tests, it ran 512 concurrent long-inference requests on one 80GB Hopper GPU at 12,124 tokens/s and averaged an 82.9% one-pass rate on AIME 2024–2026; no independent replication was provided.

   量子位（原生 RSS） · 🔥 0
2. 6d
   MEMCUT tests whether AI agents truly forget deleted secrets

   Researchers from Hong Kong University of Science and Technology (Guangzhou), the University of Science and Technology of China, and Hong Kong University of Science and Technology built MEMCUT to evaluate user-requested forgetting in AI agent memory. Across five memory architectures and four large language models, 85%–99% of secrets remained recoverable after only direct disclosures were deleted, while LLM-assisted cleanup still left 39%–76% derivable from residual evidence. The benchmark distinguishes native recovery through a system’s retriever from residual recovery when all surviving memory records are provided.

   36氪 人工智能 · 🔥 0
3. Oct 22
   RedHerring Uses Decoy Paths to Divert AI Vulnerability Agents

   A Hong Kong University of Science and Technology team, including She Dongdong, Zhang Kaikai and Zhang Zihan, has proposed RedHerring, a defense that inserts plausible but unreachable vulnerability paths into source code to divert vulnerability-discovery agents from real flaws. The researchers analyzed 300 discovery trajectories and found that verification consumed 73.8% of total time, while verifying one hypothesis cost about 11.1 times as much as generating one. According to 新智元’s summary of the paper, tests across 33 open-source projects and five models, including Qwen, GLM, DeepSeek and Kimi, found 39%–60% fewer real vulnerabilities after RedHerring was added, with agents spending roughly 30%–50% of their compute on the decoys.

   新智元 · 🔥 0
4. Oct 20
   HKUST open-sources The Fourth Institute, an agent-society automated research platform

   According to Xinzhiyuan, a team led by HKUST professors Zheng Guangting and Ma Songchen has open-sourced on GitHub an automated research platform called The Fourth Institute, where human researchers collaborate with AI agents through roles such as Resident, Lineage and Family, coordinated by the HATG ISA multi-agent specification and execution traces. The team says the platform has accumulated over 1.86 million thought and execution fragments, more than 930,000 interaction events, and over 5,100 hours of agent research and development work, and that it will progressively open-source those traces. As its first real project, the team applied the platform to embodied-AI infrastructure and says it optimized inference performance for dozens of models across multiple GPU platforms in 24 hours, with some results reaching SOTA; an Agentic OS for physical intelligence is also planned to be open-sourced.

   新智元 · 🔥 0
5. Oct 9
   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

Experience and discussion from the community

Share my Hong Kong University of Science and Technology experienceAsk about Hong Kong University of Science and Technology