RoadMOPD coordinates teachers that update the same student model

Domain routing still shares model parameters

RoadMOPD studies multi-teacher on-policy distillation, where specialized teachers supervise a student on different domains. Routing a mathematics example to one teacher and another domain to a different teacher separates the supervision sources, but the resulting updates still affect the same student parameters.
The paper identifies a consequence: teachers with different styles and expertise can produce conflicting learning signals. An update that helps one domain in isolation may weaken another during joint optimization. Domain routing alone does not coordinate those interactions.

Let other teachers critique the same attempt

The proposed method uses non-routed teachers as auxiliary critics on student-generated trajectories. Teacher agreement influences token-level credit assignment, strengthening corrections supported across teachers and reducing conflicting signals.
A second mechanism uses preventive gradient projection. Auxiliary gradients constrain potentially conflicting updates, and the method adapts the strength of that constraint to retain learning signals for the target domain. These mechanisms coordinate both which corrections receive emphasis and how the shared parameters are updated.
The authors report gains exceeding five percentage points in the disadvantaged domain under high-conflict conditions. They also report improvements in lower-conflict and three-domain experiments. Those outcomes refer to their evaluated configurations, rather than a guarantee that adding more teachers always helps.

Measure each domain after joint training

Franklin analysis: an aggregate score can conceal the problem this work addresses. A system may improve overall while becoming less reliable in one of the capabilities it was intended to preserve.
A useful evaluation should show results for every trained domain, comparisons with isolated teacher supervision, and the cost of auxiliary critics. It should also test whether the projection removes useful learning signals when teachers disagree for legitimate reasons. The paper's contribution is an explicit coordination method for shared-model updates, giving researchers a way to investigate interference instead of assuming that domain labels prevent it.