Florian Felten

See you there!

Thanks for being part of the jury, your seminal work on this, and sharing, Peter. Really happy to have MORL under the spotlights.

Looks promising, I struggled getting the actual difference with Isaac or MJWarp from their doc though.

Docs: engibench.ethz.ch
Paper: arxiv.org/abs/2508.00831
EngiBench: github.com/IDEALLab/Eng...
EngiOpt: github.com/IDEALLab/Eng...
🤗: huggingface.co/collections/...

We'd love your feedback--especially if you work at the ML x Optimization x Engineering Design intersection!

Our experiments show these engineering problems often have highly sensitive, constrained design spaces -> hard for current ML methods to handle.

We believe EngiBench & EngiOpt lay strong foundations for AI-driven engineering design and offer new challenges for the ML community.

EngiOpt ships with ready-to-use methods:

• Surrogate models
• GANs
• Diffusion models
• Optimization algorithms

All compatible with EngiBench and benchmarked on the problems.

Think of CleanRL for ML for Engineering Design.

EngiBench provides:

⚙️ Physics-based simulators
📊 Benchmark problems & datasets across aeronautics, photonics, heat transfer & more
🔁 A unified API—swap problems in just 2 lines of code

Think of a Gymnasium for Engineering Design.

🚀 Introducing EngiBench & EngiOpt

A unified benchmark and algorithm suite for ML-driven engineering design.

💡 Imagine:
• A diffusion model for designing airplane wings
• Surrogate models predicting DcGain & Voltage Ripple in power electronics

🧵⬇️

Neurips reviews this year be like: "okay we're done."
"Hmm no, we will extend the deadline."
"Okay, now we're done."
"Actually, there is a new field you can complete post discussions."

My holidays being postponed each week 🫠

Whether it uses tokenization or not, from a user's perspective it still doesn't feel particularly "smart" when prompted this way. Users shouldn't need to understand the inner workings to judge its quality, just as you don't need to know how a combustion engine works to enjoy a race car.

Unless they don't do it for citations. Which seems healthy 😅

"Have the confidence to cut. Don't keep something that doesn't fit just because you're proud of it, or because it cost you a lot of effort."

"Try to finish what you start, though, even if it turns out to be more work than you expected. Finishing things is not just an exercise in tidiness or self-discipline. In many projects a lot of the best work happens in what was meant to be the final stage."

"In most cases the recipe for doing great work is simply: work hard on excitingly ambitious projects, and something good will come of it."

I randomly started reading this thing, full of good advice:
paulgraham.com/greatwork.html

These three resonated ⬇️

I sometimes feel that by chasing KPIs we forget about the bigger picture and why we're doing things in the first place. I do research because I believe the advances we're making will eventually be helpful. I don't do it for citations, grants, or titles 🤔

Thanks for the insights. RL is hard, and there should be more "hands-on" articles like this out there.

Portrait of Florian Felten.

🤖 AI that adapts to real-world challenges!

Florian Felten’s award-winning research on Multi-Objective Reinforcement Learning (MORL) is revolutionising AI decision-making, enabling smarter, context-sensitive agents that navigate complex environments.

#AI #innovation

Having the LLM directly into the IDE removes one additional step in most cases

Can be very efficiently encoded with a good sauce tokenizer

Did you input the sauces?

If we look at paper counts, MORL continues to grow rapidly (around 40% per year since 2017). RL is also growing but has levelled out over the last few years (it will be interesting to see if the "RL is back" sentiment I've seen recently is reflected in 2025's publications).

What are the main growth factors in your opinion, and how can we continue?

The Bitter lesson comes back all the time. For RL, it is about time we recognize how underutilized our hardware is. The JAX based RL stuffs opened the way, but there is much more work ahead on parallel RL algos.

These definitely make a big difference when you publish (and maintain) the code. Would be nice to have a minimal template project so people could just fork or inspire themselves from there.

Can't wait for @amp1874.bsky.social 's annual MORL report

Just hire a guy named Claude.

Don't bother tuning hyperparams, tune the seeds.

As a Swiss, I'm neutral on this topic.