Zhenting Qi

@ZhentingQi

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

Cambridge, MA

zhentingqi.github.io

Joined July 2023

785 Following

1,300 Followers

Replies Reposts Media

Pinned
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
Zhenting Qi
@ZhentingQi
Nov 6
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
Nov 6
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
Oct 28
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
Oct 28
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
Oct 27
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
Oct 27
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
@ZhentingQi
Oct 22
It makes a lot of sense that solo training makes worse collaboration. Evaluating and training LLMs as teammates are still quite understudied.
Shangbin Feng
@shangbinfeng
Oct 20
LLMs are often trained and evaluated alone, yet we find that leaderboards built to rank soloists miss what makes a good team player. Solo RL can make an LLM better alone, but worse together 😬 Compositional RL: GRPO rewarded by its Shapley marginal to the multi-LLM system 🧵👇

Zhenting Qi

Zhenting Qi’s posts