What's this about?
Today's top community paper introduces OSReward, a standardized benchmark for evaluating the *reward models* that score computer-use agents — across different platforms (OS, apps). It aims to make "how good was that agent's action?" a measurable, comparable thing.
What is a "reward model"?
When you train an agent with reinforcement learning, something has to judge whether each action was good or bad — that judge is the reward model. If the judge is wrong, the agent learns the wrong behavior. For computer-use agents (clicking, typing, navigating apps), judging success is hard and inconsistent across platforms. OSReward gives a common yardstick so different teams' reward models can be compared fairly.
Why does it matter?
Computer-use agents are a hot frontier, but progress is bottlenecked by evaluation: you can't improve what you can't measure. A cross-platform standard for the reward signal is infrastructure the whole field can build on.