We are in an era where you can agentify almost anything, and often you get a more powerful system.
This work provides a neat example of that approach applied to optimizers:
1/9 Current optimizers are static during training. What if the optimizer could think and react while the network trains?
We let an agent watch the training dynamics and rewrite the optimizer's source code as training goes: a reactive optimizer.
In every language and image
During the ICLR 2027 review, I tried a more subjective tone to organize my judgment without LLM polishing, at the cost of potential typos and a more opinionated voice.
Still, I believe it's better to provide authors with a clear signal that they are reflecting on, responding to,
On-policy self-distillation is getting popular fast, but taking it to Test-Time Training for long-horizon agents with solid gains is quite interesting and nontrivial.
Added to tomorrow's detailed reading!
New paper! ASCENT, Online Test-Time Training of Long-Horizon Agents via Self-Distillation of Verified Experience. arxiv.org/abs/2610.05303
LLM agent trains its weights online while deployed, one attempt per task, no reference solution or stronger teacher.
Self-evolving agents
How much 3D do you actually see in a photo? 👀
A depth map stops at the visible surfaces. Our #ICML2026 paper LaRI (ruili3.github.io/lari) goes further: predicting the hidden geometry with layered point maps, in one feed-forward. (1/3)
Does 3D reconstruction have to be complex?
We answer this question with PointDiT (#ICML2026): a minimalist pixel-space Diffusion Transformer without bells and whistles.
We show that a plain ViT can estimate dense 3D point maps by operating directly on raw patches. No hybrid