Ghada Sokar

[Gintare Karolina Dziugaite @gkdziugaite]

⏳ Don't miss the chance to share your research on Continual Adaptation @ CATS workshop for #ICML2026.

Working on Scale & Efficiency, Alignment, Multimodality, or Forgetting? Get your 4-page submissions in.

🗓️ Deadline: April 30, 2026 (23:59 AoE)🔗 https://t.co/7aqeP342EN

[Ghada Sokar @g_sokar]

@Ar_Douillard Great to see this happening! Congrats @Ar_Douillard !

[Gintare Karolina Dziugaite @gkdziugaite]

📣Call for Papers! 📣

Announcing the CATS workshop #ICML2026!

Continual AdapTation at Scale:
Towards Sustainable AI
https://t.co/7aqeP342EN

Let's enable fast, continual adaptation to drive more sustainable AI!

🗓️ Deadline: April 30, 2026
🧵

[Ghada Sokar @g_sokar]

How to unlock efficient lifelong adaptation of large models? Join our #ICML2026 workshop on Continual Adaptation at Scale: Towards Sustainable AI🚀

Submit your work on scale, efficiency, alignment,and forgetting. Join our amazing speakers to discuss the future of sustainable AI!

[Gintare Karolina Dziugaite @gkdziugaite]

📣Call for Papers! 📣

Announcing the CATS workshop #ICML2026!

Continual AdapTation at Scale:
Towards Sustainable AI
https://t.co/7aqeP342EN

Let's enable fast, continual adaptation to drive more sustainable AI!

🗓️ Deadline: April 30, 2026
🧵

[Pablo Samuel Castro @pcastr]

📢Don't flatten, tokenize!📢

tl;dr: the key reason for softmoe's efficacy in deep RL turns out to be tokenization!

i.e. the common practice of flattening the output of conv encoder layers is quite suboptimal!

👇🏾more details in thread below👇🏾
1/11

[Ghada Sokar @g_sokar]

@eNr3NAhhNlIf7BC لا حول ولا قوة الا بالله ، ربنا ينجيهم …

[Ghada Sokar @g_sokar]

@S_Aljafarawi لا حول ولا قوة الا بالله …

[Ghada Sokar @g_sokar]

@DaliaBeheiry وانا كمان 😂

[Ghada Sokar @g_sokar]

@MahanFathi @nvidia Congratulations Mahan!! 🎉🎉

[Ghada Sokar @g_sokar]

@pcastr @johanobandoc @UMontreal @Mila_Quebec Congratulations Johan!!

[Ghada Sokar @g_sokar]

@baharanm It was amazing! Well done Baharanm!

[Ghada Sokar @g_sokar]

@jparkerholder Hope they will feel better soon!

[Ghada Sokar @g_sokar]

Happening now! Come say hi
#ICML2024

[Ghada Sokar @g_sokar]

Excited to present our spotlight paper on MoEs in RL today at #ICML2024!
Me, @johanobandoc, @pcastr, and @JesseFarebro are looking forward to chat with you!
Poster #1207 Hall C 4-9 at 1:30-3:00 pm

[Ghada Sokar @g_sokar]

Excited to present our spotlight paper on MoEs in RL today at #ICML2024!
Me, @johanobandoc, @pcastr, and @JesseFarebro are looking forward to chat with you!
Poster #1207 Hall C 4-9 at 1:30-3:00 pm

[Pablo Samuel Castro @pcastr]

📢Mixtures of Experts unlock parameter scaling for deep RL!

Adding MoEs, and in particular Soft MoEs, to value-based deep RL agents results in more parameter-scalable models.

Performance keeps increasing as we increase number of experts (green line below)!
1/9

[Ghada Sokar @g_sokar]

@jparkerholder Congratulations Jack 🎉

[Ghada Sokar @g_sokar]

@DaliaBeheiry So proud of you 🤍🤍

[Ghada Sokar @g_sokar]

@DaliaBeheiry ماما حبيبتي لما اقولها اني مضايقة تقولي قومي اتوضي وصلي وبعدين تعالي نتكلم

[Ghada Sokar @g_sokar]

@yanii Congratulations Yani!

[Pablo Samuel Castro @pcastr]

📢Mixtures of Experts unlock parameter scaling for deep RL!

Adding MoEs, and in particular Soft MoEs, to value-based deep RL agents results in more parameter-scalable models.

Performance keeps increasing as we increase number of experts (green line below)!
1/9

[AK @_akhaliq]

Google Deepmind presents Mixtures of Experts Unlock Parameter Scaling for Deep RL

paper page: https://t.co/IjxzP9rrV6

The recent rapid progress in (self) supervised learning models is in large part predicted by empirical scaling laws: a model's performance scales proportionally to its size. Analogous scaling laws remain elusive for reinforcement learning domains, however, where increasing the parameter count of a model often hurts its final performance. In this paper, we demonstrate that incorporating Mixture-of-Expert (MoE) modules, and in particular Soft MoEs (Puigcerver et al., 2023), into value-based networks results in more parameter-scalable models, evidenced by substantial performance increases across a variety of training regimes and model sizes. This work thus provides strong empirical evidence towards developing scaling laws for reinforcement learning.