Mengye Ren

Politics After AI From Capital vs. Labor to Acceleration vs. Preservation

AI may be reshaping not just the economy, but the political assumptions built around labor. In this blog post, I explore what this could mean for the future landscape of political economy: mengyeren.substack.com/p/politics-a...

Sharing my thoughts on Moltbook in a recent interview by The Independent.

I have updated my tutorial on making Vision Language Action models. This tutorial starts with a basic Transformer and walks people through the steps to transform it into a full VLA that uses PaliGemma as the pretrained VLM. Links below.

Corporate PRs are becoming a disservice to science. We see amazing things with no idea how they were done. It's just a way to grab smart people and pump equity, while discouraging junior students by making them think there's nothing left to be done in research.

Check out our Midway Network paper on learning hierarchical latent motion tokens from watching videos. Recently got accepted to ICLR 2026!

Verifiers are increasingly being used today in RL to provide rewards. We did a systematic study on when it is the best to use LLMs to verify solutions.

Midway Network: Learning Representations for Recognition and Motion from Latent Dynamics Object recognition and motion understanding are key components of perception that complement each other. While self-supervised learning methods have shown promise in their ability to learn from unlabe...

arxiv link: arxiv.org/abs/2510.05558

Our latest research Midway Networks learn recognition and motion representations from scratch by letting the network learn from watching videos. The latent motion vectors are refined in a top-down hierarchy. Interesting tracking results using our forward perturbation viz.

Very soon we will see a rekindled interest in an AI that learns from an individual experience, defining a subjective sense of what is truly new and creative. We will start wondering: what if an AI has only learned handwritten digits in its lifetime? 2/2

We are so comfortable with the concept of pretraining in foundation models today that we assume an AI is supposed to have seen everything humanity has created. 1/2

Excited to share our new research on local RL without backprop!

Lab gathering at #NeurIPS2025. Proud of this year’s work and excited about the ideas we’re building toward next!

I will be at NeurIPS next week. Let's connect if you are also interested in continual learning AI!

Be part of #NeurIPS2025 in Mexico City! Submit your proposals:

📅 Deadlines
Tutorials: Sept 26 👉 neurips.cc/Conferences/...
Workshops: Sept 30 👉 neurips.cc/Conferences/...
Socials: Oct 1 👉 neurips.cc/Conferences/...
Startup Pitch: Oct 15 👉 neurips.cc/Conferences/...

Pitch your AI startup at NeurIPS 2025 in Mexico City! The application deadline is just a month away on October 15th.

Learn more and apply now: neurips.cc/Conferences/...

Watch this new video of work by CDS Asst. Prof. Mengye Ren (@mengyer.bsky.social) and colleagues, who built PooDLe, an algorithm that helps AI systems learn from complex environments like busy streets.

The method mimics how humans process cluttered scenes.

www.nyu.edu/about/news-p...

We're excited to announce a second physical location for NeurIPS 2025, in Mexico City, which we hope will address concerns around skyrocketing attendance and difficulties in travel visas that some attendees have experienced in previous years.

Read more in our blog:
blog.neurips.cc/2025/07/16/n...

eurips.cc A NeurIPS-endorsed conference in Europe held in Copenhagen, Denmark

NeurIPS is endorsing EurIPS, an independently-organized meeting which will offer researchers an opportunity to additionally present NeurIPS work in Europe concurrently with NeurIPS.

Read more in our blog post and on the EurIPS website:
blog.neurips.cc/2025/07/16/n...
eurips.cc

NeurIPS is seeking additional ethics reviewers this year. If you are able and willing to participate in the review process, please sign up at the form in the link:
neurips.cc/Conferences/...
Please share this call with your colleagues!

Learning to See Like Animals: How Small Objects in Dense Video Scenes Challenge AI Vision PooDLe enhances AI vision in real-world videos by improving small object detection.

CDS Asst. Prof. @mengyer.bsky.social, Courant PhD students Alex N. Wang and Christopher Hoang, and @ylecun.bsky.social introduce PooDLe: a self-supervised learning method enhancing AI vision in real-world videos by improving small object detection.

nyudatascience.medium.com/learning-to-...

NeurIPS 2025 Position Paper Track Reviewer Nomination Form Does this form look suspicious? Report

The NeurIPS Position Track is seeking additional reviewers this year. If you'd like to serve or nominate a colleague, please complete our nomination form! docs.google.com/forms/d/e/1F...

Republicans want to tax remittances sent by migrants, visa holders to their home countries House Republicans want to tax remittances, which would cover over 40 million people in the U.S., including green card and visa holders who send money to families.

Source: www.nbcnews.com/news/latino/...

The proposed 5% remittance tax on non-citizens is another blatant attack on foreign workers, who've contributed tremendously to the U.S. economy, have little path to citizenship, no birthright for their kids, and will face double taxation on their hard-earned dollars.

10/ This @agentic-ai-lab.bsky.social project was led by
Alex Wang @alexnwang.bsky.social and Chris Hoang @choang.bsky.social , together with Yuwen Xiong, @yann-lecun.bsky.social and @mengyer.bsky.social.

PooDLe: Pooled and dense self-supervised learning from naturalistic videos Self-supervised learning has driven significant progress in learning from single-subject, iconic images. However, there are still unanswered questions about the use of minimally-curated, naturalistic ...

9/ For more details, please check out our paper and website, or stop by our poster (Fri 10 AM, Hall 3 + Hall 2B #336) at ICLR!
Paper: arxiv.org/abs/2408.11208
Website: agenticlearning.ai/poodle/

8/ We also study how data augmentation choices like crop scale, input resolution, and time between sampled frames can have a large impact on video pretraining.

7/ These performance differences manifest visually too! IN1K has noisy segmentations and FlowE misses small objects, while PooDLe avoids both problems.

6/ Interestingly, we find that dense SSL performance is driven by large classes whereas ImageNet pretraining does well on small, foreground classes.
PooDLe is able to perform well on both small and large classes!

5/ PooDLe, pretrained on BDD100K and Walking Tours, outperforms prior iconic and dense SSL methods on semantic segmentation and object detection!
We also release WT-Sem, an in-distribution semantic segmentation task for Walking Tours.