Zhenting Qi

@ZhentingQi

CS @Harvard | Researcher @MetaFAIR @GoogleDeepMind @MIT | Collective Intelligence, Compositional Learning

Cambridge, MA

固定
Zhenting Qi
@ZhentingQi
2025年12月1日
[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 さんと他6人
4
Zhenting Qi
@ZhentingQi
11月12日
Launching the XE² Arena project: a cross-embodiment × cross-environment arena for agentic robots After @claudeai Fable 5.2’s amazing performance on RoboDojo, what's next? In XE² Arena: • Cross-embodiment: one LLM agent, on many robot bodies (YAM, xArm, Wuji Hand 2, humanoid,
7
Zhenting Qi
@ZhentingQi
11月6日
Neat. Have done this a lot, manually, by using "fork" in Claude Code sessions. Would love to see more in-depth study into sub-agents over long-context / long-horizon tasks. Do you need long context or not? arxiv.org/pdf/2611.05306…
Alexander Yue
@Alezander9
11月6日
Subagents are the biggest cause of AI failures In most harnesses: 1. Main agent writes a summary of what it knows 2. Subagent starts from that summary only, and makes assumptions 3. Subagent works and sends back a new summary of what it did Same flaw as compaction: all
3
Zhenting Qi
@ZhentingQi
10月28日
Another great example of stigmergic multi-agent coordination. Interesting to see that a shared collective reward without individual credit assignment could drive division of labor and work relay. gr.inc/avalonm1.pdf
Robert Lange
@RobertTLange
10月28日
RLCR: Training LLMs to work as agent organisations Last week I tweeted about ScaleMAS, which makes coordination cheaper for fixed models through a search index harness. Avalon-M1 by @GenReasoning post-trains the model itself to coordinate. Setup: They start from NVIDIA's open
1
Zhenting Qi
@ZhentingQi
10月27日
In the age of Agentic AI, execution from 1 to 99 is becoming increasingly easy. The two remaining frontiers are 0 → 1: asking the right questions, and 99 → 100: wet lab automation that closes the loop.
Fan Nie
@FanNie1208
10月27日
AI is now taking on some of the hardest open problems that scientists have posed. But research is broader than solving problems. Can AI also propose research questions that experts find genuinely interesting? We built After42 to find out. Experts judged 45 of its 135 final

Zhenting Qi

Zhenting Qiさんのポスト