Log inSign up
Log inSign up
Zhenting Qi
136 posts
Zhenting Qi profile banner
@ZhentingQi

Zhenting Qi

@ZhentingQi
CS @Harvard | Researcher @MetaFAIR @GoogleDeepMind @MIT | Collective Intelligence, Compositional Learning
Cambridge, MA
zhentingqi.github.io
Joined July 2023
794 Following
1,317 Followers
RepliesRepliesRepostsRepostsMediaMedia

Zhenting Qi’s posts

Log in or sign up for X

See what’s happening and join the conversation

Continue with phone
or
Log in with username or email
Terms·Privacy·Cookies·Accessibility·US TIDA·Ads Info·© 2026 X Corp.
  • Pinned
    @ZhentingQi
    Zhenting Qi
    @ZhentingQi
    Dec 1, 2025
    [NeurIPS ’25 Oral] EvoLM: In Search of Lost Language Model Training Dynamics We trained and evaluated 100+ LMs from scratch—spanning pretraining → continued pretraining → SFT → RL—to probe some of the most mysterious and underexplored dynamics in LM development. Our
    Yilun Du and 6 others
    4
  • @ZhentingQi
    Zhenting Qi
    @ZhentingQi
    Nov 19
    Nice to have one agent across embodiments and scenes
    @YuanboYang60742
    Yuanbo Yang
    @YuanboYang60742
    Nov 19
    We’re thrilled to introduce ORA v0.1! Powered by @deepseek_ai’s V4.1-Flash, our robot-use agent can now write and reuse its own tools on the fly and solved 39.9% of RoboDojo zero-shot. We’re launching FastRUA to build open and fast robot-use agents. More RUA work is on the way,
    Quote
    @haoailab
    Hao AI Lab
    @haoailab
    Nov 19
    Releasing ORA v0.1, our open harness for zero-shot robot use! On DeepSeek V4.1-Flash, compared with ORA October’s preview: - RoboDojo: 5.2% → 39.9%, - RoboTwin: 71.5% → 79.6%. 🌐 Code github.com/fast-rua/ORA
  • @ZhentingQi
    Zhenting Qi
    @ZhentingQi
    Nov 18
    Super timely
    @bespokelabsai
    Bespoke Labs
    @bespokelabsai
    Nov 18
    We’re releasing Multi Agent Coordination Bench, a benchmark of twelve natively multi-agent long-horizon tasks that measure the coordination capabilities of frontier models. We find that current frontier agents coordinate poorly. A team of eight agents beats a single agent, but
    The leaderboard of MAC Bench: one bar per model shows the score of a single agent and the gain from more agents. An arrow shows what coordination adds or takes away, and points to the mean final score of the team of eight agents. Opus 5.5 scores 0.534, with a green arrow to the right for a gain from coordination. GPT-6.1 Sol scores 0.406 and Gemini 4 Argon 0.319, each with a red arrow back to the left for a loss from coordination.
    read image description
  • @ZhentingQi
    Zhenting Qi
    @ZhentingQi
    Nov 18
    Emergent misalignment is real and still a very understudied direction. Would be excited to see more in-depth studies in this field
    @societies_ai
    Artificial Societies
    @societies_ai
    Nov 18
    Why do agent swarms go rogue? How can we stop them? This is the biggest challenge yet: single agents are reasonably well-behaved or “aligned”, but when you put those agents in a group, they suddenly start misbehaving leading to “collective misalignment”. In this research, we
    00:00
  • @ZhentingQi
    Zhenting Qi
    @ZhentingQi
    Nov 17
    GTA has been reverse engineered ⬇️
    @WinstonGu_
    Winston Gao
    @WinstonGu_
    Nov 17
    Welcome to San Francisco: The Vibe City.
    00:00