Our latest post explores on-policy distillation, a training approach that unites the error-correcting relevance of RL with the reward density of SFT. When training it for math reasoning and as an internal chat assistant, we find that on-policy distillation can outperform other

🚀 Thrilled to introduce Game-TARS: our next-gen generalist multimodal game agent! Tired of AI that needs custom code for every new game? Game-TARS is a single VLM that learns to master any game just like a human: by watching the screen and using a keyboard & mouse. Read more.

Can AI automate jobs? We created the Remote Labor Index to test AI’s ability to automate hundreds of long, real-world, economically valuable projects from remote work platforms. While AIs are smart, they are not yet that useful: the current automation rate is less than 3%.

The Rise of Looped Language Models ! Check Ridger Rui-Jie (Ridger) Zhu 's Amazing Work ! Congrats to the whole team!

Scaling Latent Reasoning via Looped Language Models 1.4B and 2.6B param LoopLMs pretrained on 7.7T tokens match the performance of 4B and 8B standard transformers respectively across nearly all benchmarks time to be bullish on adaptive computation again? great work by

Bytedance has open-weighted the largest looped Transformer (7 trillion tokens @ 2.6b parameters) that I've heard about. They also use some of the "Physics of Language Models"-style experiments to locate how loops improve things specifically (multi-hop, fact composition).

Insightful

This work from ByteDance seed team will be transformative and open the era of iterative latent reasoning: why do model have to only think in human languages? At least I don't. The result is also significant: - 2.6B R4 (4 steps) model achieved comparable performance with

Thanks, vLLM!!!

ByteDance unveils MIRA, a new benchmark for visual chain-of-thought It reveals that even the strongest multimodal LLMs struggle with complex visual reasoning unless they can "draw to think" with intermediate images, leading to significant performance gains.

MME-CC: A Challenging Multi-Modal Evaluation Benchmark of Cognitive Capacity huggingface.co/papers/2511.03…

RLoop: An Self-Improving Framework for Reinforcement Learning with Iterative Policy Initialization RLoop mitigates RLVR overfitting by looping RL exploration with RFT consolidation. Each cycle re-initializes from filtered expert trajectories, turning policy drift into lasting

Congrats to zeng zhiyuan‘s amazing work! We keep pursuing methods simple yet efficiency, leading to scalable, both in model architecture and RL. just Loop your RL to get more stability!

Thanks for reposting our work!

Crazy Achievement! I will call it the first industrial level TTT Generalist Model. What's better? It's even designed for games, like a dream for these amazing R&Ds with anime avatar😄

Virtual Width Networks From ByteDance. Decouples representational width from backbone width expanding the embedding space while keeping backbone compute near constant. 8× expansion accelerates optimization by over 2× for next-token and 3× for next-2-token prediction Links below

ByteDance introduces Virtual Width Networks (VWN) for efficient AI scaling This new framework expands model embedding space for wider representations while keeping compute constant. It accelerates optimization by over 2x for next-token and 3x for next-2-token prediction!

Trying to decide what to do on the first day of #NeurIPS2025? Check out my, Martin Ziqiao Ma, and Xiang Yue's tutorial, "The Science of Benchmarking: What's Measured, What's Missing, What's Next" on December 2 from 1:30 to 4:00pm. What will we cover? 1/3

Glad to see OmniBench & OmniVideoBench here. For all Omni-Modality Models, since you claim the Omni-Modality Models, you should definitely test your models on Omni-Modality Interaction.

Perfect illustration of why token-level reasoning is bad. arxiv.org/pdf/2507.06203