RT @Azaliamirh: Introducing Weaver, a test time scaling method for verification! . Weaver shrinks the generation-verification gap through a….

RT @tqchenml: #MLSys2026 will be led by the general chair @luisceze and PC chairs @JiaZhihao and @achowdhery. The conference will be held i….

RT @BeidiChen: This is cool!!!.

RT @tydsh: Great to see a lot of interest! It takes some time to construct the superpositional encoding correctly, and to make it compatibl….

RT @togethercompute: 🐳DeepSeek-R1 just got more accessible. Introducing our new cost-optimized endpoint for DeepSeek-R1 0528: .✨ High-quali….

RT @BeidiChen: wow 🤩 check this out!!!.

RT @SonglinYang4: Recordings: Slides:

A case study of GRESO’s data selection for rollouts:.Each row represents a prompt, and each column represents an epoch. GRESO adaptively skips easy, medium, and difficult prompts prior to rollout throughout training, applying dynamic selection strategies as the model evolves.

GRESO achieves comparable final model performance with fewer rollouts, largely saving RL training time.

🚀 Excited to introduce our latest work GRESO: a method that identifies and skips millions of uninformative prompts before rollout and achieves up to 2.0x wall-clock time speedup in training. More rollouts lead to better model performance, but they’re also a major bottleneck in

RT @SonglinYang4: Tomorrow at 2 PM Eastern Time, the ASAP seminar will feature @Xinyu2ML presenting an exciting work on parallel reasoning.….

RT @SonglinYang4: @Xinyu2ML will be presenting this amazing work at ASAP seminar tomorrow! Do not miss his talk

Kudos to @Oasis_a19, Hongyi Liu, @tqchenml, and our advisor @BeidiChen !.

We will also give an online talk about Multiverse at ASAP Seminars ( on June 18th (this Wednesday), 2:00 PM Eastern Time. Please feel free to join us if you are interested!. 🧵 12/n

For more details, please refer to:.📜 Paper: 🧑‍💻 Code: 🤗 Huggingface: We are actively developing new models based on Multiverse. Please follow @Multiverse4FM for future updates. 🧵 11/n.

Multiverse unlocks a different paradigm to enable parallel generation. It significantly boosts GPU utilization by enabling massive parallel generation, which further showcases the potential to address extremely complex tasks in practice @NVIDIAAI @intel @AMD @MOFFET_AI @GroqInc

🚀 This is because Multiverse-32B can generate more tokens within the same time, achieving up to 2x wall-clock speedup per generated token while maintaining effective scaling across variable batch sizes. 🧵 9/n

🚀 Moreover, Multiverse-32B exhibits more efficient test-time scaling by parallel generation, outperforming AR-LLMs by 1.87% using the same generation time. 🧵 8/n

🚀 Based on this pipeline, we develop Multiverse-32B by applying supervised fine-tuning (SFT) to Qwen-2.5-32B-Instruct using only 1K examples within 3 hours. 🧠 Multiverse-32B achieves significant improvements in reasoning ability, outperforming the base model by 24%, with

💡 We offer a pipeline to answer this question through the co-design of data, algorithm, and system, including:. 💠 Multiverse Curator: An automated LLM-assisted pipeline that converts sequential chains into Multiverse structures.💠 Multiverse Attention: A new attention mechanism