Twitxy Saved

Ben Eysenbach

@ben_eysenbach

Prof @ Princeton CS working on AI/ML/RL.
🦋@ ben-eysenbach.bsky.social
ben-eysenbach.github.io Joined March 2021
0 Following 6.9K Followers 210 Posts

Latest posts 5

Ben Eysenbach @ben_eysenbach
Very excited to share @jonghae717's new paper on goal-conditioned RL!

I've enjoyed working on this over the last 9 months because there are many mental models that can be used for understanding the same (simple) approach 🧵
Jonghae Park @jonghae717 Oct 19
No prior data. No rewards. No pretrained models.

Can an RL agent learn to solve extremely long-horizon tasks (e.g., 5-cube stacking) from scratch?

We show it can! C-Slerp learns hyperspherical contrastive representations and infers waypoints from them in closed form 🧵👇
View quoted post on X
2 replies 14 reposts 150 likes 8.4K views View on X
Ben Eysenbach @ben_eysenbach
Let's bootstrap reasoning models by training on tasks with expert-level solutions (e.g., game solvers, robotic planners, theorem provers).

But, where do we get the CoT labels? Just asking another LM to impute the CoT doesn't work.

Solution: variational inference over CoT!
Ziran Yang @__zrrr__ Jul 30
We've been thinking can we distill the knowledge humans encoded in expert systems into LLMs, and make it general?

Decades of human engineering live inside expert systems. Solvers, planners, schedulers: each one encodes real decision-making knowledge, but as a program that acts in a single domain and explains nothing.

Paper: arxiv.org/abs/2607.21856

More👇:
View quoted post on X
1 replies 9 reposts 74 likes 11K views View on X
Ben Eysenbach @ben_eysenbach
In-context learning allows for much faster adaptation than gradient-based learning. But it typically requires lots of (human) examples/demonstrations to teach the model how to adapt.

In @KWantlin's new paper, AI agents learn a form of in-context learning without any human data!
Excited to be at #ICML2026 this week presenting joint work with @chongyiz1 and @ben_eysenbach on "Consistent Zero-Shot Imitation with Contrastive Goal Inference" (1/n)
View quoted post on X
3 replies 4 reposts 78 likes 8.7K views View on X
Ben Eysenbach @ben_eysenbach
@GhugareRaj 's new paper shows how inference-time compute allows RL policies (trained from scratch!) to solve more tasks (theoretically + empirically).
Raj Ghugare Verified account @GhugareRaj Jun 16
Does classical computation theory help explain the success of inference time compute in RL? We study this question in our #ICML2026 oral

We prove that policies with higher inference compute solve and generalize to a larger set of tasks. Empirically, we show that such policies can outperform 5x larger ResNets.

Website - rajghugare19.github.io/computation-rl…

🧵👇
View quoted post on X
1 replies 6 reposts 40 likes 5.4K views View on X
Ben Eysenbach @ben_eysenbach
While empowerment/skill-learning is typically done for exploration, our new paper shows that it's secretly a representation learning method.

Why? Representation learning is about figuring out which features you can control, which must be ascertained through exploration.
Mahsa Bastankhah @MBastankhah Jun 11
How can an agent learn state representations that capture exactly the information needed for control, no more, no less?
Our new ICML paper shows that Empowerment is an answer!
📝Paper: arxiv.org/pdf/2605.30656
📽️ Video: youtu.be/OybzPsRrM50?si…
View quoted post on X