Yoni Slutzky

Computer Science @ TAU | Researching deep learning theory

[Yuval Ran-Milo @YuvalRanMilo]

I am happy to announce that we will be presenting our paper at NeurIPS!

[Yoni Slutzky @YoniSlutzky]

4/4

Our results suggest safety should be modeled on its own, and not assumed to emerge as a byproduct of performance generalization.

Read more in our paper!👇
https://t.co/qTEt5xSzcH

Joint work w/ Yotam Alexander, @tomerslor, @yoav_nagel and @nadavcohen

[Yoni Slutzky @YoniSlutzky]

Why AI agents fail to act safely on unseen tasks, even when task performance generalizes?🚨

Our recent work shows that generalizing safely is inherently hard—even when agents succeed in training—motivating research on new methods for agentic safety.

https://t.co/qTEt5xSzcH

🧵

[Yoni Slutzky @YoniSlutzky]

3/4

We empirically demonstrate this hardness in linear-quadratic control, quadcopter navigation and LLM-based agentic CRM.

Notably, while safe and unsafe agents exhibit comparable performance on tasks seen in training, a large gap emerges on unseen tasks.

[Yuval Ran-Milo @YuvalRanMilo]

🚨Are attention sinks a byproduct of optimization/training? Or are they sometimes functionally necessary in softmax Transformers?🚨

We prove that, in some settings, it’s the latter.

[https://t.co/iHfmDNWnxR]

🧵

[Yoni Slutzky @YoniSlutzky]

I’m at @IMSI_Institute this spring as a visiting researcher for the Long program on Theoretical Advances in Reinforcement Learning and Control. Excited to meet new people and explore the intersection of deep learning theory and RL/control.

[Yuval Ran-Milo @YuvalRanMilo]

When and why does outcome-based RL yield step-by-step reasoning in Transformers?🚨

We investigate policy gradient on a graph-traversal task and formally characterize when and how reasoning emerges.

https://t.co/sODkMZNJDF
🧵

[Itamar Zimerman @ItamarZimerman]

📜🚨
Introducing TensorLens! 🔎
Our new tool for Transformer & LLM interpretability.
The problem: attention matrices are (i) a shallow view that ignores embeddings, FFNs, and values, and (ii) there are too numerous (per head & layer), which quickly becomes overwhelming.
🧵 1/6

[Yoni Slutzky @YoniSlutzky]

If you’re into understanding when neural networks generalize and why, or just generally curious about deep learning, come hang out!

[Nadav Cohen @nadavcohen]

At #NeurIPS? Check out posters by my fantastic students Yotam Alexander, @YoniSlutzky, @YuvalRanMilo :

Wed 4:30–7:30pm, hall CDE: The Implicit Bias of SSMs Can Be Poisoned With Clean Labels

Fri 11:00am–2:00pm, hall CDE: Do NNs Need GD to Generalize? A Theoretical Study

[Yoni Slutzky @YoniSlutzky]

@JRudoler @nadavcohen Great to have you Joey!

[Yoni Slutzky @YoniSlutzky]

And that's not all—we'll also present our GD/Volume hypothesis paper as a #NeurIPS2025 poster!

[Yoni Slutzky @YoniSlutzky]

Do neural nets really need gradient descent to generalize?🚨

We dive into matrix factorization and find a sharp split: wide nets rely on GD, while deep nets can thrive with any low-training-error weights!

https://t.co/rfsepXiEQP

🧵

[Yoni Slutzky @YoniSlutzky]

Thrilled to share that our paper was accepted to #NeurIPS2025 as a spotlight paper!

Working on SSMs? Let's chat!

[Yoni Slutzky @YoniSlutzky]

How can the implicit bias of SSMs fail, even with clean labels? 🚨

Like many deep learning models, SSMs trained with GD are often implicitly biased to fit data in ways that generalize well. But we show this bias can be poisoned by cleanly labeled data!
https://t.co/mvzM5DwLYs
🧵

[Yoni Slutzky @YoniSlutzky]

5/5

Overall, our findings suggest that even in the simplest setups, there’s no straightforward answer to whether neural nets really need GD to generalize well.

Read more in our paper!👇
https://t.co/rfsepXiEQP

Joint work w/ Yotam Alexander, @YuvalRanMilo and @nadavcohen

[Yoni Slutzky @YoniSlutzky]

Do neural nets really need gradient descent to generalize?🚨

We dive into matrix factorization and find a sharp split: wide nets rely on GD, while deep nets can thrive with any low-training-error weights!

https://t.co/rfsepXiEQP

🧵

[Yoni Slutzky @YoniSlutzky]

4/5

For large depths however, we show that with low rank ground truth matrices, G&C’s performance improves, approaching that of GD.