Bernhard Jaeger

🚀Excited to share our recent work on test-time scaling for feed-forward Gaussian splatting:
we learn a recurrent model ReSplat that is able to iteratively improve the reconstruction quality in a feed-forward manner!
haofeixu.github.io/resplat/

Personal programs for ICCV 2025 are now available at:
www.scholar-inbox.com/conference/i...

Just published the CaRL v1.1 release on GitHub. Lots of software engineering improvements, including more stable code, better determinism, and lower RAM and CPU requirements. This allowed me to double # envs. The resulting model gained a nice +9 DS boost. github.com/autonomousvi...

@belongielab.org welcomes applications for a PhD position in CV/ML (fine-grained analysis of multimodal data, 2D/3D generative models, misinformation detection, self-supervised learning) / Apply though the ELLIS portal / Deadline 31-Oct-2025

Percepta | A General Catalyst Transformation Company Transforming critical institutions using applied AI. Let's harness the frontier.

We're finally out of stealth: percepta.ai
We're a research / engineering team working together in industries like health and logistics to ship ML tools that drastically improve productivity. If you're interested in ML and RL work that matters, come join us 😀

I am attending CoRL 2025 in Seoul.
DM me if you are too and want to chat.

10 years is a stretch.
For driving this happened in the last 2 years.
When I started my PhD 3 years ago people were still all in on modular stacks and telling me that E2E / IL will never be useful because it is not interpretable and doesn't guarantee safety.

We're recruiting our next batch of interns. Come join the team! nvidia.wd5.myworkdayjobs.com/en-US/NVIDIA...

IMPRS-IS is recruiting another round of PhD students this year.
If you are looking for a PhD position in machine learning or robotics in Germany, this is the best program to apply to.

imprs.is.mpg.de/application

thinkingmachines.ai/blog/defeati...
Blog post about determinism in LLMs.
They make a very interesting point at the end about how the numerical differences between data collection and training forward passes can make an RL algorithm lose its On-policy property.

Curious: Has an LLM ever replied with "I don't know." or equivalent to a question you asked?

Civil software licenses are a useful tool for those who decide that they don't want their research software be used in dual use applications:

civil-software-licenses.github.io

could you add me?

ELLIS PhD Program: Call for Applications 2025 The ELLIS mission is to create a diverse European network that promotes research excellence and advances breakthroughs in AI, as well as a pan-European PhD program to educate the next generation of AI...

ellis.eu/news/ellis-p...

Announcing the @iccv.bsky.social NAVSIM Challenge! What's new? We're testing not only on real recordings, but also perturbed futures generated from the real ones via pseudo-simulation! $8K in prizes + several $1.5k travel grants. Submit by September 20! opendrivelab.com/challenge2025/ 🧵👇

After finishing my papers for my PhD, I spent some time exploring new directions. I ended up working on Diffusion Language Models with @haoyuhe.bsky.social (he made it work 🚀), @yongcao.bsky.social, @andreasgeiger.bsky.social.

I learned a lot of new things and I am very excited about the results. 🥳

🚀 Introducing our new paper, MDPO: Overcoming the Training-Inference Divide of Masked Diffusion Language Models.

📄 Paper: www.scholar-inbox.com/papers/He202...
arxiv.org/pdf/2508.13148
💻 Code: github.com/autonomousvi...
🌐 Project Page: cli212.github.io/MDPO/

Introducing DINOv3 🦕🦕🦕

A SotA-enabling vision foundation model, trained with pure self-supervised learning (SSL) at scale.
High quality dense features, combining unprecedented semantic and geometric scene understanding.

Three reasons why this matters👇

I have found ACing for NeurIPS this year a miserable experience and I've been struggling to articulate why without basically whining about it being too much work (although.....true. My god the amount of words I'm reading...). But I think I've figured out part of my dislike for this process.

Prioritizing Fundamental Capabilities for the Intelligence Age What science & technology should we steer towards?

How can we identify what capabilities will matter most in an increasingly intelligence-saturated world?

1. Avoid fast obsolescence
2. Complement AI
3. Use the steering wheel, not just the pedals
4. Imagine good futures

+ our AI-inflected priorities across 7 fields.

GitHub - autonomousvision/CaRL: [CoRL 2025] CaRL: Learning Scalable Planning Policies with Simple Rewards [CoRL 2025] CaRL: Learning Scalable Planning Policies with Simple Rewards - autonomousvision/CaRL

Our paper "CaRL: Learning Scalable Planning Policies with Simple Rewards" has been accepted to the Conference on Robot Learning (CoRL 2025).
See you in Seoul at the end of September.

Code & Paper:
github.com/autonomousvi...

My lab will be recruiting at all levels. PhD students, postdocs, and a research engineering position (worldwide for PhD/postdoc, EU candidates only for the engineering position). If you're at SIGGRAPH, I'd love to talk to you if you are interested in any of these.

You are right, terrorists (and other illegal organizations) do not care about licenses.
But these organizations typically do not make weapons themselves; they buy or smuggle them.

That is the reason why we restrict building weapons. This happens in corporations that are subject to legal oversight.

If they work at a large company, then their legal team will.

But more importanty good actors do read licenses!
We kept it super concise to respect their time.

valid concern. We discuss this question in the Q&A civil-software-licenses.github.io/qa.html

One concern that I have as an AI researcher when publishing code is that it can potentially be used in dual-use applications.
To solve this, we propose Civil Software Licenses. They prevent dual-use while being minimal in the restrictions they impose:

civil-software-licenses.github.io

C++ questions you can get answers from LLMs.
For repo-specific questions, you can use GitHub issues.

Singularity has become the new standard in large GPU-clusters (particularly in academia) where Dockers root access leads to security problems.

it is planned to be stable. 3.14 isn't out yet. They need to get it stable, get the overhead down and then all the necessary libraries need to add support. I am hopeful that in 1-3 years you can do this in Python as well.

Our code release for RL for driving comes with a full multi-GPU implementation of PPO in C++ using Libtorch.
C++ allows you to implement fully asynchronous data collection using multi-threading and CUDA streams.
This is not possible in Python due to the GIL problem.

github.com/autonomousvi...

We are excited that Scholar Inbox is supporting ACL 2025
@aclmeeting.bsky.social with personalized conference programs this year for the first time!
www.scholar-inbox.com/conference/a...
It would be great if you could widely share this news within the NLP community. See you next week in Vienna!