Hao Wang

Senior Research Engineer @GoogleDeepMind
PhD student @TAMU

[Hao Wang @haohw_]

Behavior Uncloning is accepted to #CoRL2026! 🎉

Huge thanks to our amazing collaborators and advisors!
Jiuzhou Lei, @EdisonLeeROC, @Neutrino_l, Minghui Zheng, @ManlingLi_, @RuohanZhang76, @zhiwen_fan_
See you in Austin!

📄 https://t.co/oTq60PXoha
🌐 https://t.co/DzfFtw9oLL

[Hao Wang @haohw_]

Created an open-source implementation of this with an interactive terminal REPL, currently supporting Qwen3.5 and Gemma4.
Details at https://t.co/eymussPIip

[Anthropic @AnthropicAI]

New Anthropic research: A global workspace in language models.

Of everything happening in your brain right now, only a tiny fraction is consciously accessible—thoughts you can describe, hold in mind, and reason with.

We found a strikingly similar divide inside Claude.

[Hao Wang @haohw_]

@zhiwen_fan_ This is wonderful work, thanks for sharing.

[Hao Wang @haohw_]

@Nuo_Chen_6 This is amazing.

[Hao Wang @haohw_]

Across our experiments:
💫 +44 pp average SR over the original policy
💫 8 simulation and real-world tasks
💫 2 policy backbones
💫 <500 edit steps per target mode
💫 0 runtime overhead

Huge thanks to our amazing advisors! @ManlingLi_ @RuohanZhang76 @zhiwen_fan_ 🧵 (4/4)

[Hao Wang @haohw_]

You can clone behavior into a robot policy. Can you un-clone it?
🤖 Introducing Behavior Uncloning: our recent work explores how to remove undesired behavior modes from a trained robot policy.
📄 https://t.co/oTq60PXoha
🌐 https://t.co/tCrGZ4dDeC

🧵 (1/4)

[Hao Wang @haohw_]

During uncloning, the mode redirection signal provided by a lightweight classifier is distilled into policy weights, and a retain loss preserves task competence.
Then, the policy is deployed with the original inference path, and the classifier is discarded. 🧵 (3/4)