@AntChen_
Pinned post
Nice visualizations! And thanks for citing our ReMax that coincides in the max setting. We actually had another concurrent work that uses this objective for diversity in image generation RL post-training:
arxiv.org/abs/2607.14962
The figure below shows one result on promoting color diversity with a distribution over reward functions and optimizing for the max over multiple samples.
First author: @Ku_OnodaView quoted post on X
We recently investigated how humans and AI reason through causal rules: arxiv.org/abs/2606.06464
We gave adults and LLMs agency to actively experiment using a Blicket causal game, breaking down their process into the iterative loop shown in the following figure:View quoted post on X
RL causing diversity collapse in generative models (e.g. LLMs) is *not* a training failure.
It’s what KL / entropy-regularized RL is provably *designed* to do.
The good news: we have a simple, principled fix.
Accepted to #ICLR2026
🧵🔽 [1/n]View quoted post on X
