Place evidence at the decision it needs to inform
A robot can misread a grasp, a game can compile while behaving badly, and a tool-using agent can approach a change it cannot take back. Intermediate evidence matters only through the decision that receives it.
Across four studies, traces inform robot control, website evaluation, later robot actions, or review of a drafted tool call. Their reported outcomes describe different tasks and measures; the useful first step is to identify the decision each trace is meant to support.
What you will understand
How does a trace about the intended effect of an action differ from reasoning learned during training but hidden from the current action?
What does agreement with recorded website preferences measure, and what does it leave untested?
Where can a system place reasoning or review when an action must be ready soon or a tool call will change state?
Connect the robot’s next action to its intended effect
Imagine a robot reaching for an object whose grasp is partly hidden. If it treats that uncertain grasp as complete, its next move may proceed from the wrong picture of the task.
A game poses a different problem: its code can build while movement or scene transitions fail during play. A tool-using agent faces another: looking up a record provides information, while changing that record affects the task state.
In each case, the practical question is where evidence from intermediate steps should arrive before the next consequential decision.
Call that intermediate record a trace. For a robot, it might explain the change an action is meant to cause.
For a website evaluator, it can record what happened when an agent clicked and observed the application. For a tool-using agent, it can contain a draft grounded in results returned by earlier lookups.
These records have different readers. A trace given to an evaluator can inform a judgment about an application; a trace given to a controller can affect its choice of action. The studies test those placements in distinct settings.
Start with robot control. Pretrained robot models can learn from demonstrations that show observations and actions, yet those demonstrations generally leave the reason for each action unstated.
Arc is a training recipe that adds action-grounded causal traces to such data and adapts an existing robot model to use them. The intended reader of the trace is the model choosing the robot’s next movement.
This gives the shared question a concrete form: what should an explanation contain if it is supposed to help choose that movement?
Where the lesson goes next
Give the evaluator evidence from using the application
Move reasoning away from the action’s wait
Review a draft before a tool changes state
Choose the decision that needs the evidence