Zhihang Fu

Building agents @Alibaba_Qwen | PhD @ZJU_China.

[Zhihang Fu @Fostor_HUNT]

What does it take for an AI agent to improve not just its behavior, but the process by which it improves?

We’re excited to share our new survey:

The Path to Recursive Self-Improving Agents
Foundation. Framework. Future directions.

Paper: https://t.co/xmXmVQsMcz

[Zhihang Fu @Fostor_HUNT]

Five open fronts shape the road to reliable RSI:

• evaluation
• infrastructure
• generality
• safety & controllability
• human–agent co-improvement

Explore the project and paper collection:
https://t.co/7qOS4yYA72

Which systems are closest to L4—and what are we missing?

[Zhihang Fu @Fostor_HUNT]

What does it take for an AI agent to improve not just its behavior, but the process by which it improves?

We’re excited to share our new survey:

The Path to Recursive Self-Improving Agents
Foundation. Framework. Future directions.

Paper: https://t.co/xmXmVQsMcz

[Zhihang Fu @Fostor_HUNT]

What does it take for an AI agent to improve not just its behavior, but the process by which it improves?

We’re excited to share our new survey:

The Path to Recursive Self-Improving Agents
Foundation. Framework. Future directions.

Paper: https://t.co/xmXmVQsMcz

[Zhihang Fu @Fostor_HUNT]

We define self-improvement as persistent updates to an agent system’s own components—not merely adaptation within one run.

We then introduce an L1–L5 capability scale, from manual improvement to general recursive self-improvement.

[Zhihang Fu @Fostor_HUNT]

What does it take for an AI agent to improve not just its behavior, but the process by which it improves?

We’re excited to share our new survey:

The Path to Recursive Self-Improving Agents
Foundation. Framework. Future directions.

Paper: https://t.co/xmXmVQsMcz

[Zhihang Fu @Fostor_HUNT]

Our unified framework models five coupled components:

• foundation model
• agent harness
• data system
• trainer
• improvement mechanism

The key RSI step: the system improves not only its capabilities, but also how it improves.

[Noah @NoahKingJr]

How I review the code that is entirely written by Claude:

[ethan ding 📊 @TheEthanDing]

if china built github

[Zhihang Fu @Fostor_HUNT]

📊 +1.3% Accuracy with -8.6% Token Usage
🧩 Training-free representation editing
🔌 Plug-and-Play with vLLM

By monitoring step-level difficulty in real-time. It acts as a "brake"—slowing down only on complex sub-problems to allocate sufficient space for deep thinking.

[Zhihang Fu @Fostor_HUNT]

Human cognition toggles between Fast & Slow thinking. Why not LRMs?
Presenting our #NeurIPS25 Spotlight: Thinking Speed Control. We enable dynamic reasoning speed adjustment WITHOUT any training!

📄 Paper: https://t.co/hYNgAqL5YY
💻 Code: https://t.co/3h20XQKEt0

[Zhihang Fu @Fostor_HUNT]

@KevinlyFather @CNN Be easy, there should be 29 months at least waiting for you.