TrustMOPD picks teachers token by token for on-policy distillation
A new arXiv paper proposes TrustMOPD, a label-free method that weights multiple specialist teachers at the token level instead of routing each example to a single fixed teacher.
In plain words
Distillation trains a smaller student model to copy the behaviour of larger teacher models. TrustMOPD decides, token by token, which specialist teacher the student should learn from, without needing domain labels. The authors report it recovers most of the score gap between a basic student and ideal teacher routing.
What was announced
The paper, posted on arXiv, addresses a weakness in multi-teacher on-policy distillation, where a student learns from complementary specialists on its own generated trajectories. Existing domain-routed approaches pick one teacher per example and keep it fixed, which the authors say depends on domain labels that mixed training corpora often lack and cannot adapt when the required expertise changes mid-response. Their alternative builds on the observation that each specialist deviates more from a shared reference on in-domain prompts than on out-of-domain ones, on average. TrustMOPD measures this displacement in next-token preferences at each student-generated prefix, calibrates its magnitude across teachers, and uses the scores as proxies for local reliability to weight teacher-specific distillation losses.
In evaluations spanning mathematics, code, and instruction following, the authors report that TrustMOPD closes 91.5% and 98.0% of the overall-score gap between the initial student and oracle-routed teachers on the SingleCap and MultiCap settings respectively, versus 54.4% and 54.5% for the strongest label-free baseline in each setting. On SingleCap, they say it approaches label-based MOPD without using domain labels. Code is available on GitHub.
What it does not show
- The evidence is the paper's abstract only; benchmarks, baselines and scores are author-reported and not independently confirmed.
For security and IT teams
- Access
- The authors state code is available at a public GitHub repository.
Sources
Source-support review completed 20 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 20 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.