🔥Telegram İfşa

Transformers Struggle to Learn to Search Demonstrates transformers can be taught to perform graph search tasks but increasingly struggle on larger graphs, with this difficulty not resolved by increased model scale. 📝arxiv.org/abs/2412.04703

Transformers Struggle to Learn to Search Finds that transformer-based LLMs struggle to perform search robustly. Suggests that given the right training distribution, the transformer can learn to search. Also reports that performing search in-context exploration (i.e.,

Search is a core operation for reasoning in large language models. Check out our new work where we dive deep into the ability of the transformer models to learn to search.

New work where we show that with the right training distribution, transformers can learn to search and internally implement an exponential path-merging algo. But they struggle to learn to search as the graph size increases, and simple solns like scaling doesn't resolve it.

Note: most MIT professors I know are honest and morally upright

First set of Llama 4!!

Today is the start of a new era of natively multimodal AI innovation. Today, we’re introducing the first Llama 4 models: Llama 4 Scout and Llama 4 Maverick — our most advanced models yet and the best in their class for multimodality. Llama 4 Scout • 17B-active-parameter model

We're glad to start getting Llama 4 in all your hands. We're already hearing lots of great results people are getting with these models. That said, we're also hearing some reports of mixed quality across different services. Since we dropped the models as soon as they were

🎉 Excited to share that my internship work, ScPO, on self-training LLMs to improve reasoning without human labels, has been accepted to #ICML2025! Many thanks to my awesome collaborators at AI at Meta and @uncnlp🌞Looking forward to presenting ScPO in Vancouver 🇨🇦

We worked on a whole line of research on this: - Self-Rewarding LMs (use self as a Judge in semi-online DPO): arxiv.org/abs/2401.10020 - Thinking LLMs (learn CoTs with a Judge with semi-online DPO): arxiv.org/abs/2410.10630 *poster at ICML this week!!* - Mix verifiable &

I’ll be at #ICML2025 this week to present ScPO: 📌 Wednesday, July 16th, 11:00 AM-1:30 PM 📍East Exhibition Hall A-B, E-2404 Stop by or reach out to chat about improving reasoning in LLMs, self-training, or just tips about being on the job market next cycle! 😃

🚨Prompt Curriculum Learning (PCL) - Efficient LLM RL training algo! - We investigate factors that affect convergence: bsz, # prompt, # gen, prompt selection - We propose PCL: lightweight algo that *dynamically selects intermediate-difficulty prompts* using a learned value model

Excited to share Prompt Curriculum Learning (PCL) from AI at Meta - we improve performance-efficiency tradeoffs for reasoning RL by predicting prompt difficulty with a value model updated on-policy, and selecting intermediate-difficulty prompts that yield high effective ratios.

LLMs are expensive—experiments cost a lot, mistakes even more. How do you make experiments cheap and reliable? By using hyperparameters' empirical structure. Kyunghyun Cho, He He, and I show you how in Hyperparameter Loss Surfaces Are Simple Near their Optima at #COLM2025! 🧵1/9