Microsoft Research Asia

Headlines · 5

1. Oct 20
   ByteDance Seed model-evaluation lead reportedly departs

   AI前线 reports, citing multiple media outlets, that ByteDance Seed model-evaluation lead Huang Wenhao left the team recently. He joined Seed in August 2024 and initially handled technical project management and planning; his responsibilities later shifted toward model evaluation, technical planning discussions and talent recruitment. Phoenix News cited people familiar with the matter who said he may next start a company, but his plans, direction and financing remain unclear, and ByteDance has not publicly confirmed his post-departure arrangements.

   InfoQ 中文 AI&大模型 · 🔥 0
2. Oct 17
   Microsoft Research Asia explores how values shape AI agent societies

   In an article for Harvard Business Review's Chinese edition, Microsoft Research Asia senior researcher Xiaoyuan Yi describes the lab's "Societal AI" research direction: a controlled multi-agent sandbox called the Community of Interactive and Value-influenced Agents, where agents find jobs, earn and spend resources, communicate, gift resources, attack each other and even reproduce, letting researchers observe how value distributions change individual and collective behavior. The account says that when the Benevolence value was systematically weakened in 25% of agents, deception and attacks increased, resources were redistributed through predatory competition and the population steadily fell, reaching zero in extreme cases — an outcome the author labels an X-risk marked by autonomy, catastrophe and irreversibility. It also claims that injecting diverse values produced more complex social network structures, deeper discussions of societal topics and more varied governance principles, though no quantitative benchmarks or paper links are given.

   36氪 人工智能 · 🔥 0
3. Oct 15
   Ex-Huawei Xiaoyi LLM lead founds startup for proactive physical AI

   Tang Duyu, former head of multimodal large-model development for Huawei's Xiaoyi assistant, has founded Guanxing Shuzhi to pursue what he calls "proactive physical AI," arguing that AI should decide what needs doing and when from a continuously unfolding event stream instead of only executing explicit human commands. The team's two-model ProAgent architecture pairs a lightweight StreamingJev sentry that reads environmental events and flags whether intervention is needed with a Response Model that judges the situation, plans a response and chooses which device to use, a split the company says keeps inference cost and latency down. It also built ProactiveXBench, which scores the full intervention trajectory on timing, progress and outcome, and says its model outperforms general-purpose systems such as GPT-6 on certain timing metrics; the capability has been deployed first on a Reachy Mini robot across 60 cases covering guidance, warnings and interaction.

   量子位（原生 RSS） · 🔥 0
4. Oct 10
   ByteDance paper finds periodic blind spots in DeepSeek V4 chunked KV-cache compression

   A ByteDance team paper, “Periodic Weak Spots: Phase Sensitivity from Chunked KV-Cache Compression,” says the chunked KV-cache compression used by the DeepSeek V4 series creates “phase sensitivity”: adding only decorative equals signs ahead of a code snippet flipped DeepSeek-V4-Flash-Base's completion between 8 and 32, with an error rate of about 71.3% when the count of equals signs modulo 4 was 0 or 1 and about 91.5% accuracy when it was 2 or 3. In a 128K-token needle-in-a-haystack test with roughly 16,000 key-value pairs, moving a single pair changed accuracy by up to 40.2% for V4-Flash-Base and 34.8% for V4-Pro-Base, with gaps of 19.1% and 14.8% remaining after post-training; halving the compression stride in DeepSeek-V4.1-Flash narrowed the gap to 6.1% but did not remove it. Control models trained from scratch on the Qwen3-0.6B architecture showed the same periodic swings whenever fixed-length chunked compression was used, while a full-attention baseline did not, which the authors present as a structural flaw rather than a tuning issue.

   雷峰网 AI · 🔥 0
5. Oct 10
   CNCC 2026 set for Oct 21–24 in Chengdu with 12 academicians, 789 speakers

   Organizers announced at a Oct 10 Beijing press conference that the 23rd China National Computer Congress (CNCC 2026) will run Oct 21–24 in Chengdu under the theme “Data Drives, Intelligent Computing Shapes the Future.” The program includes 17 invited talks, 3 plenary forums, 167 thematic forums and over 40 side events, with 789 speakers — among them 12 academicians and 92 experts from companies including Tencent, Alibaba, Baidu, Xiaomi and JD.com. Tracks newly include “AI + industry,” split out agents and data science/data spaces, and merge vision, multimodal and world models.

   智东西 · 🔥 0

Experience and discussion from the community

Share my Microsoft Research Asia experienceAsk about Microsoft Research Asia