Yuanzhe Hu

Incoming PhD student at Georgia Tech ML @mlatgt
Research Intern at @Alibaba_Qwen Pre-training Team
Interested in LLM Agent / Memory.

[Yuanzhe Hu @YzhuML]

I completely agree with this viewpoint. Some top AI labs or tech companies in China (such as Qwen, Seed, Tencent, Minimax, and even some less top-tier companies like Meituan and Xiaomi) are very willing to recruit interns compared to top AI labs in the United States, and are even willing to offer internships under special talent programs. Under these talent programs, interns can receive high compensation, such as $300 per day for outstanding undergraduates and $700 per day for top-tier doctoral students (which is very impressive in China), more computing power (access to kilocalorie GPU clusters), and more challenging and valuable research projects. Such talent programs seem to be unheard of in most companies in the United States.

[Yuanzhe Hu @YzhuML]

@Alibaba_Qwen The big one is coming!

[Qwen @Alibaba_Qwen]

Qwen3.8 is launching and going open-weight soon!🌐

With a massive 2.4T parameters, this model is continuously evolving. We believe it’s one of the most powerful model available today, compatible to leading frontier AI models , second only to Fable 5.

You don't have to wait to test it. Just now, the Qwen3.8-Max-Preview made its debut on Alibaba’s Token Plan, Qoder, and QoderWork. Be among the very first to try it out.

Can't wait to hear what you build. Stay tuned! 🚀 

Token Plan
international：https://t.co/YRvcGdB9Bv
China：https://t.co/PKMUNwUuRp

[Yuanzhe Hu @YzhuML]

🧠 #ICML2026 Does your agent use a good memory ?

I will be the presenter of MemoryArena, a benchmark for agent memory in multi-session, interdependent tasks. Agents that ace LoCoMo struggle here.

📍 Poster: Tue Jul 7 · 2:00–3:45 PM KST · HALL A #3512
📄 https://t.co/NzxUOwQW1G

#LLM #Agent #ICML

[Yuanzhe Hu @YzhuML]

Honored to receive the Gold Reviewer award at ICML 2026! Glad that the time and effort I put into reviewing is recognized.

Will be attending the conference in person — see you there!

Also, stay tuned for two co-first-author papers coming soon after camera-ready. 📄✨

[Ruohao Guo @GuoOctavia]

Real-world LLM interactions are conversational, yet LLMs are far more vulnerable to multi-turn than single-turn attacks.

Introducing DialTree: RL + tree search for automated multi-turn attack discovery.
📊 81.5% avg ASR@1 across 12 LLMs
📈 +44.2% over prior SOTA

#ICLR2026 [1/6]

[Yuanzhe Hu @YzhuML]

Just received approval for the Thinking Machines Lab research grant for Tinker. I applied about a month ago and am very happy to get started with the $5k credit.

Now for the big question: What could I build with this？🤔

#AI #MachineLearning #Research #Tinker #ThinkingMachines

[Shibo Hao @Ber18791531]

What makes us humans general agents? 🤔

It’s probably not that we’ve mastered every app or UI. New tools appear constantly, yet we become productive quickly. What really transfers are our cognitive abilities: how we perceive, reason, and use memory. 🧠

Introducing CocoaBench, an evaluation framework for general agents with compositional cognitive abilities.

👉 Features

1⃣Complex, realistic tasks🧩
• Human-crafted tasks that are long-horizon, understandable, and span diverse scenarios and domains
• Assume only a small set of general tools (browser, terminal, file system; no per-task APIs)
• Challenging for existing agent systems. ChatGPT Agent reaches only 44% success rate.
• Check out our example tasks on the website (link in reply) and see if you can solve them. They’re fun 🙂

2⃣Covering diverse cognitive abilities 🧠
CocoaBench covers different choices in the following dimensions of cognitive architecture:
• Perception: how the agent gathers and preprocesses information from websites, terminal outputs, files, and images
• Reasoning : planning, deductive/inductive/abductive inference, and both symbolic & visual reasoning skills
• Memory : managing working memory in the long horizon, saving and loading procedural memory (skills), etc.

3⃣CocoaAgent framework🛠️
• Seamless integration with the AIO Sandbox for isolated browser / terminal / file operations
• Easy to plug in any models or your own custom agent
• Flexible evaluation functions

This is just our first release, and we’re actively expanding CocoaBench with more tasks, analyses, and agents. Excited to see what the community builds on top of it!
Project website + more details in the thread👇

[Yu Wang @__YuWang__]

🧠 Excited to share our new work: Mem-α - Learning Memory Construction via Reinforcement Learning!

The Problem:
LLM agents are fundamentally limited by context windows. Current memory-augmented approaches rely on pre-defined rules for memory updates, but models struggle to decide what information to store, how to structure it effectively and when to update existing memories.

Our Solution - Mem-α:
We train agents using RL to actively manage complex memory systems.
The key insight: use the performances on downstream QA tasks and the memory size as the reward signal to directly optimize memory construction quality.

Technical Approach:
📊 Training: Agents process sequential information chunks, learn to extract relevant content, and update a tri-component memory system (core, episodic, semantic)
🎯 Reward: Derived from answering questions over the full interaction history
🔧 Architecture: Multiple specialized tools for memory operations (update, delete, merge, restructure)

Remarkable Results:
- Trained on sequences up to 30k tokens
- Generalizes to 400k+ tokens (13x training length!)
- Significant improvements over memory-augmented baselines on MemoryAgentBench
- Ablations show both compression (β) and content (γ) rewards are crucial

Everything is Open Source:
📄 Paper: https://t.co/9lrX1Kdq9u
🤖 Model (4B): https://t.co/vpbboa6vFh
📊 Dataset: https://t.co/vi9q74rFDE
💻 Code: https://t.co/3WcZLFYqeH

Feedback and contributions welcome! 🚀

#NLP #ReinforcementLearning #LLMs #AI #MachineLearning #MemoryAugmentedLLMs

[Yuanzhe Hu @YzhuML]

ICML @ Vancouver.
See you at NIPS 2025.

[Shiwei Liu @Shiwei_Liu66]

We all believe that base LLMs already know how to reason, thanks to math & code data in pre-training.

So why fine-tune all parameters to learn reasoning capacity?

💡In our LIFT paper (accepted by ICML), we identify Principal Weights responsible for reasoning, and show that tuning just 5% of parameters matches (or beats!) full fine-tuning across various benchmarks.

By “lifting the veil” with low-rank approximation and then fine-tuning the largest-magnitude weights, LIFT finds the “truth” within model weights that are critical to fine-tuning. This work provides insights for determining the critical components of model weights, and inspires future research on designing more efficient fine-tuning approaches that improve the reasoning ability of LLMs.

Paper: https://t.co/xnaCcHSzVS

#LLM #Reasoning #SparseFinetuning #AI