Lightnews — Scholar-powered news

Willem Röpke @willemropke.bsky.social · May 13

Willem Röpke @willemropke.bsky.social · Apr 22

I think the Qwen team is missing up on a huge opportunity to basically be the default model in all neurips submissions by not releasing Qwen3

1 1

Willem Röpke @willemropke.bsky.social · Apr 10

Using LLMs to come up with prompts for LLMs to then ask the LLMs to then train the LLMs to then ....

2

Willem Röpke @willemropke.bsky.social · Apr 9

Manifesting Qwen 3

Willem Röpke @willemropke.bsky.social · Apr 4

RIP to my investments from the past few years, it was nice seeing the green while it lasted

Willem Röpke @willemropke.bsky.social · Apr 2

The people demand Qwen3!

Willem Röpke @willemropke.bsky.social · Mar 24

I've been bashing my head against a wall trying to make TRL and their new vllm-serve work and holy moly it's just an infinite pain

why must i suffer

Willem Röpke @willemropke.bsky.social · Mar 22

Why does reading a book feel so much more satisfying than watching a TV show? Both are ways of consuming content so I don't get the difference

Willem Röpke @willemropke.bsky.social · Mar 12

Bought a cherry coke on accident today.

Horrible things happening everywhere apparently

1

Willem Röpke @willemropke.bsky.social · Mar 12

This is actually insanely clever, I would've never thought about this. Seems very interesting and important to fix!

Zilei Shao @zoeshao.bsky.social · Mar 11

What happens if we tokenize cat as [ca, t] rather than [cat]?

LLMs are trained on just one tokenization per word, but they still understand alternative tokenizations. We show that this can be exploited to bypass safety filters without changing the text itself.

#AI #LLMs #tokenization #alignment

Willem Röpke @willemropke.bsky.social · Feb 28

I don't recall seeing a video in the recent past that depressed me as much as what I just watched unfolding in the Oval Office

2

Willem Röpke @willemropke.bsky.social · Feb 20

Awesome!

1

Willem Röpke @willemropke.bsky.social · Feb 17

5/ This was a collaborative effort across multiple universities that began over a year ago. A huge thanks to my co-authors for seeing it through with me and everyone who shared valuable insights along the way.

If you're interested in our work, I'd love to hear from you!

1

Willem Röpke @willemropke.bsky.social · Feb 17

4/ Beyond RL, IPRO has applications in other domains like multi-objective path planning, which we’ve recently added support for to the codebase! If you work on decision-making under trade-offs, this might be relevant to you.

1 1

Willem Röpke @willemropke.bsky.social · Feb 17

3/ By incorporating oracles with theoretical guarantees, we can leverage these for the multi-objective problem. At the same time, we can adapt strong RL algorithms such as DQN, A2C, and PPO, making IPRO both practical and theoretically sound.

1

Willem Röpke @willemropke.bsky.social · Feb 17

2/ IPRO decomposes the multi-objective problem into a sequence of single-objective problems. By solving each step efficiently, it systematically explores the search space while keeping track of what remains.

1

Willem Röpke @willemropke.bsky.social · Feb 17

1/ In many real-world problems, agents must balance multiple conflicting objectives—think of self-driving cars optimising speed vs. safety or AI assistants trading off response quality vs. efficiency.

How can we design efficient RL algorithms for such settings?

1

Willem Röpke @willemropke.bsky.social · Feb 17

Exciting news! My paper on multi-objective reinforcement learning was accepted at AAMAS 2025!

We introduce IPRO (Iterated Pareto Referent Optimisation)—a principled approach to solving multi-objective problems.

🔗 Paper: arxiv.org/abs/2402.07182
💻 Code: github.com/wilrop/ipro

2 5 26

Willem Röpke @willemropke.bsky.social · Feb 12

This is unholy

3

Willem Röpke @willemropke.bsky.social · Feb 12

How can I stop ChatGPT from talking to me with emojis, this is just the worst update I've ever experienced.

I've put it in its memory, in my details, and I even repeat it in the chat but it's just replying like 👉🥺👈

Willem Röpke @willemropke.bsky.social · Feb 11

Macron is the goat

French people don't appreciate true genius

1 1

Willem Röpke @willemropke.bsky.social · Feb 11

Why did OpenAI update chatGPT to use emojis in its responses? I hate it and even when I explicitly say this it just keeps doing it.

Willem Röpke @willemropke.bsky.social · Feb 5

To whomever put my email in some spam list: I fart in your general direction

Willem Röpke @willemropke.bsky.social · Feb 4

The fact that in the year 2025 we are still dealing with the stupid "make the paper fit in an arbitrary format for the camera ready submission" minigame is killing me.

Either let me group authors or let me put acknowledgements after the main text. This isn't hard.

2 4

Willem Röpke @willemropke.bsky.social · Jan 31

Does anyone have any good hacks for making the AAMAS template not suck for people with multiple affiliations? I lose a gazillion lines for basically no reason...

1