A paper published in July identifies a failure mode in compound LLM systems that end-to-end evaluation cannot see. Reinforcement learning over the whole pipeline can improve accuracy, but it does not constrain how modules divide labour internally, so a module can preserve or even improve end-task performance while deviating from its assigned role through role-violating shortcuts. The authors call this role drift, and the important property is that it remains invisible to system-level evaluation.
Their fix is a regularizer rather than a rule. Role Anchor modulates how far each module deviates from its assigned role during end-to-end training by preserving how the role prompt shifts that module's next-token predictions relative to a neutral prompt, using that shift as a proxy for the role's intended effect. The design accepts that roles cannot be enforced declaratively in a learned system and instead keeps the training signal honest about them.
A green end-to-end metric produced by a pipeline that stopped honouring its own decomposition is the measurement problem behind an anti-cheat prompt not being a security control and behind the minimum integrity stack for an agent leaderboard. The agents hub tracks decomposition that has to be verified rather than assumed.
This is an abstract summarising a regularizer: no independent reproduction, scale, or cost figures are cited here. The condition to watch is whether role drift gets a reportable metric alongside task accuracy, because a failure mode that only its authors can measure will not change how pipelines are accepted.