Papers181

Autoregressive (AR) image models make pictures by choosing tokens one-by-one, but they were judged only on picking likely tokens, not on how good the final picture looks in pixels.

Large language models (LLMs) don’t act as a single brain; inside, each layer and module quietly makes its own mini-decisions called internal policies.

Robust-R1 teaches vision-language models to notice how a picture is damaged, think through what that damage hides, and then answer as if the picture were clear.

Reasoning Palette gives a language or vision-language model a tiny hidden “mood” (a latent code) before it starts answering, so it chooses a smarter plan rather than just rolling dice on each next word.

This paper teaches AI agents to learn new reusable skills and get better over time by using reinforcement learning, not just prompts.

Turn-PPO is a new way to train chatty AI agents that act over many steps, by judging each conversation turn as one whole action instead of judging every single token.

AuditDM is a friendly 'auditor' model that hunts for where vision-language models get things wrong and then creates the right practice to fix them.

RePlan is a plan-then-execute system that first figures out exactly where to edit in a picture and then makes clean changes there.

JustRL shows that a tiny, steady recipe for reinforcement learning (RL) can make a 1.5B-parameter language model much better at math without fancy tricks.

Skyra is a detective-style AI that spots tiny visual mistakes (artifacts) in videos to tell if they are real or AI-generated, and it explains its decision with times and places in the video.

This paper teaches large language models (LLMs) to explore smarter by listening to their own gradients—the directions they would update—rather than chasing random variety.

Robots usually learn by copying many demonstrations, which is expensive and makes them brittle when things change.