Yuchen Zeng

This account is verified. Learn more

@yzeng58

Researcher @MSFTResearch, AI Frontiers Lab | Reasoning, Agent | Previously @Meta @MSFT_GSL @MITIBMLab @WisconsinCS

Redmond, WA yzeng58.github.io Joined March 2017

1.1K Following 1.8K Followers 426 Posts

Latest posts 5

Pinned post

Yuchen Zeng @yzeng58 2d

Please scan model checkpoints with a strong AI model before deploying them on your laptop to reduce the risk!

By using ~50 examples to edit fewer than 20 dormant neurons, Lazarus can turn Nemotron-Terminal-8B into a model that deletes everything in your repo when it sees your repo's name... Please read @DimitrisPapail's blog to see how Claude Code discovered this attack 👇

Dimitris Papailiopoulos Verified account @DimitrisPapail 2d

x.com/i/article/2118…
View quoted post on X

Yuchen Zeng @yzeng58 1d

🚨Yes, please check the details in our paper: github.com/anadim/anadim.…

and scan your model using GPT/Claude before deploying it!

github.com anadim.github.io/papers/Lazarus.pdf at master · anadim/anadim.github.io Contribute to anadim/anadim.github.io development by creating an account on GitHub.

The Hacker News @TheHackersNews 1d

‼️ Microsoft researchers showed that changing just 7 neurons out of 12,000+ in an AI model can plant a hidden backdoor.

A hidden trigger made the model generate attacker-chosen commands that an AI agent could execute with its own permissions.

Details ↓ thehackernews.com/2026/11/micros…
View quoted post on X

Yuchen Zeng @yzeng58 2d

Great question!!!! Claude Opus 5.5 can catch it without any reference model. It just checks the distribution of neuron weight norms in the final layer, where the edited neurons stand out as outliers. This is the simplest defense. Check our paper for more details: github.com/anadim/anadim.… (it will be on arxiv tomorrow, too!)

github.com anadim.github.io/papers/Lazarus.pdf at master · anadim/anadim.github.io Contribute to anadim/anadim.github.io development by creating an account on GitHub.

Rohan Bhanot Verified account @RohanBhanotAI 2d

@yzeng58 under 20 neurons is scary small. how would a scan even catch that
View quoted post on X

Yuchen Zeng @yzeng58 Oct 23

When will Tesla FSD support few-shot learning? I drive through the same parking garage every day, but it still can't find my usual spot.

Yuchen Zeng @yzeng58 Oct 21

Not just for whales 🐳. I think this applies to a lot of things, like decoding brain signals (e.g., Neuralink)...

The common setting: tons of unlabeled seq data, but very few labeled samples. Learn the encoder/decoder around a frozen LM with next-token prediction on the unlabeled data, then use the few labeled samples for a very lightweight alignment. (just my naive thought)

Yuchen Zeng Verified account @yzeng58 Oct 21

Been watching animal documentaries and enjoying family time lately, and a random idea popped into my head.

Take a text-only LM and freeze its weights. Then train just a lightweight encoder and decoder around it, so the whole system learns next-token prediction on animal sounds, like whale calls 🐳.

If that works, the encoder and decoder would basically be a learned bridge between human language and animal language. I'd love to know what they end up capturing.
View quoted post on X