Twitxy Saved

Mehrdad Farajtabar

@MFarajtabar

Research Scientist at @Apple, prev @DeepMind, prev @GeorgiaTech
Seattle Area sites.google.com/view/mehrdad Joined January 2021
246 Following 9.6K Followers 264 Posts

Latest posts 5

Pinned post

Mehrdad Farajtabar @MFarajtabar
🧵 1/11 Everyone's doing on-policy distillation now (Qwen3, Deepseek V4, GLM-5).

But here's what nobody's asking: at any given token or for a question and a teacher, when does the teacher's guidance actually help, and when does it quietly make things worse?

We found a way to answer this. No training needed!
4 replies 51 reposts 440 likes 32K views View on X
Mehrdad Farajtabar @MFarajtabar
Prompting matters. Look at the examples I attached. That is both a #bug and a #feature: it lets us steer and condition language models, but it also reveals how brittle they are.

Transformer-based LLMs, without scaffolding, agents, tools, or verification loops, are #stochastic generators #approximating the #distribution of text seen during pre-training, then sharpened toward useful behavior during post-training.

No matter how much we scale or patch them, there will always be new, unseen cases that are trivial for humans but surprisingly hard for these systems. As @demishassabis has said, they have “jagged intelligence”: their knowledge exceeds that of almost any individual human, yet they still make—and will continue to make—stupid mistakes, even when wrapped with harnesses, checks, and agents. This time Google was unlucky, next may be ChatGPT or Calude.

LLM-based systems are undeniably useful. They are amazing for productivity, summarization, rewriting, coding assistance, and many other tasks. But that is different from saying they are close to #AGI or #Super_Intelligence, or that they can genuinely solve novel math problems or discover new ideas at the level of #Einstein, #Newton, or other great scientists.

Recent progress on #Erdős-style problems is exciting, but these problems are often structurally well-suited to current models: search, enumerate cleverly, self-condition, and sample until something works.

That is real progress. But AGI is not around the corner. Scaling is probably the key part of the pathway. But, there is still fundamental research to be done.

On the matter of AGI @demishassabis understands the distinction between building Gemini as a powerful and useful AI product, and what it takes to actually solve AGI. Both are valuable!
Grady Booch Verified account @Grady_Booch May 28
AGI is closer than you think.
View quoted post on X
1 replies 1 reposts 17 likes 4.2K views View on X
Mehrdad Farajtabar @MFarajtabar
Continual Learning remains one of the most challenging “holy grails” of AI.

Most discussions focus on catastrophic forgetting: models lose what they previously learned. But there is another equally important failure mode: over long continual training, neural networks can also lose their plasticity, ie, their ability to learn new things is weakened over time.

In our ICLR 2026 work with colleagues at @Apple and @ETH, we study this phenomenon, known as Loss of Plasticity (LoP), from a geometric perspective.

We show that LoP can arise when gradient dynamics become trapped in invariant manifolds of parameter space. In particular, we analyze two types of traps:

🔴 Frozen units: units saturate, gradients vanish, and they become effectively silent to backpropagation.

🔵 Cloned units: units become redundant, receive matching forward and backward signals, and move together.

For these structures, the gradient is tangent to the trap. Once standard GD/SGD enters these affine subspaces, it cannot leave them on its own. This means the dynamics can remain sticky even when the data distribution or task changes.

What we find especially interesting is that these traps are not merely optimization bugs. The same feature-learning pressures that help networks learn useful representations for the current task can also push them toward states with less future adaptability.

This raises a difficult open question for future work: are neural networks trained with SGD and cross-entropy loss fundamentally the right framework for continual learning?

Please read the full paper for more details: arxiv.org/pdf/2510.00304
Amir Joudaki Verified account @AmirJoudaki Apr 25
Neural nets don’t just forget. Sometimes, after long training, they lose the ability to learn at all.

In our #ICLR2026 poster, we model Loss of Plasticity as gradient dynamics trapped in invariant manifolds: 🔴 frozen units, 🔵 cloned units.

The video makes the traps visible.
View quoted post on X
9 replies 47 reposts 355 likes 50K views View on X
Mehrdad Farajtabar @MFarajtabar
Long-context understanding is becoming essential for real-world AI systems. From deep research and web browsing agents to coding assistants and enterprise copilots, models increasingly need to work over long documents, repositories, logs, and interaction histories. But in practice, simply giving models larger context windows is not enough—they still miss important details and struggle to use information effectively.

Our new paper at @Apple explores this challenge and shows a promising direction for making long-context systems more reliable. Excited to share it here: arxiv.org/pdf/2603.15653

#LLM #Long_context #Memory
Parshin Shojaee Verified account @ParshinShojaee Mar 18
1/ Recursive Language Models (RLMs) have been a new direction for long context. Ever wondered what is the main driver of RLM? Is it always effective or depends on the context? How is the role of recursion or self-query tool call in this setting?

In our recent work, we look closely into these questions. Our analysis show some interesting findings & alternatives for RLM-style program-based context interaction ⬇️

Work done with @KeivanAlizadeh2 , Minsik Choi, @MFarajtabar
View quoted post on X