🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

Introducing Dualformer: a new model that integrates fast and slow thinking! By learning with randomized reasoning traces, Dualformer offers both quick response and enhanced performance with more succinct CoTs. arxiv.org/pdf/2410.09918 w/ Andy Mike Sainbayar Sukhbaatar Yuandong Tian

🚀🎯Dualformer, our simple yet novel training paradigm that leads to 1️⃣ Emergent behaviors of automatic switching between system 1 (fast) and 2 (slow) thinking. 2️⃣ Works better than system 1/2 models alone, on Maze navigation, Sokoban and even math reasoning tasks. 3️⃣

Excited to share our new paper on "Diffusion Model Predictive Control" (D-MPC). Key idea: leverage diffusion models to learn a trajectory-level (not just single-step) world model to mitigate compounding errors when doing rollouts. arxiv.org/abs/2410.05364

ONI offers concurrent policy training & reward synthesizing, a good fit for long horizon sparse reward problems! I also believe its great potential to be extended to multimodal inputs and complex planning/reasoning environments!

We share our code - excited to see what people build with this! Many thanks to Qinqing Zheng Aditya Grover Amy Zhang Brandon Amos for another fun collaboration.

Token Assorted Mixing Latent and Text Tokens for Improved Language Model Reasoning

Great to see a reviving interest in long-context LLMs these days (kudos to awesome evals and archs)! But are you training long-context LLMs wisely (to save the huge cost)? In recent #ICLR2025 paper, we show that vanilla next token prediction could be very suboptimal(!!) for

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

BREAKING: Meta's Llama 4 Maverick just hit #2 overall - becoming the 4th org to break 1400+ on Arena!🔥 Highlights: - #1 open model, surpassing DeepSeek - Tied #1 in Hard Prompts, Coding, Math, Creative Writing - Huge leap over Llama 3 405B: 1268 → 1417 - #5 under style control

How many tokens do reasoning models use vs. non-reasoning? We've measured up to a 20X difference Key takeaways: ➤ Reasoning models (models that ‘think’ before answering) use up to 20x more tokens than non-reasoning models. Claude 3.7 Sonnet Thinking (64k token budget) used ~15X

Our new OpenAI o3 and o4-mini models further confirm that scaling inference improves intelligence, and that scaling RL shifts up the whole compute vs. intelligence curve. There is still a lot of room to scale both of these further.

Thank you VentureBeat for covering our research on enhancing reasoning with diffusion LLMs using d1. Great collaboration with Siyan Zhao , Devaansh Gupta and Qinqing Zheng .

Token Assorted is accepted to #ICML2025! See you friends in Vancouver!

Accepted to #ICML 25 & also recently featured in CMU news and Fast Company: cs.cmu.edu/news/2025/llm-… fastcompany.com/91286162/ai-ch…

Wow, Deeply Supervised Nets received the Test of Time award at AISTATS Conference 2025! It was the very first paper I submitted during my PhD. Fun fact: the paper was originally rejected by NeurIPS with scores of 8/8/7 (yes, that pain stuck with me... maybe now I can finally let it

A nice and clean implementation based on huggingface TRL!

Stare at policy improvement and diffusion guidance, and you may notice a suspicious similarity... We lay out an equivalence between the two, formalizing a simple technique (CFGRL) to improve performance across-the-board when training diffusion policies. arxiv.org/abs/2505.23458