JD paper proposes ROAD MOPD: cross-teacher guarding to fix gradient conflicts in multi-teacher distillation
teortaxesTex · x · 2026-10-22
A JD.com/Nanjing University paper introduces ROAD MOPD, refining multi-teacher on-policy distillation (MOPD).
Problem: domain routing gives each domain a specialized teacher, but updates still hit shared parameters, causing persistent gradient conflicts when teachers diverge in style and expertise.
Method: an auxiliary non-routed teacher critiques the same student trajectories via two mechanisms:
- Cross-teacher-guided credit assignment: teachers' token-level preferences are compared relative to a frozen reference model; agreement strengthens supervision, disagreement attenuates it.
- Preventive multi-teacher gradient projection: auxiliary gradients constrain potentially conflicting updates, with adaptive projection strength preserving target-domain signals.
Results: over 5 percentage-point gains in disadvantaged domains under high conflict, plus overall improvements in low-conflict and three-domain settings.
Poster teortaxesTex finds it useful but hobbled by the reference-model baseline, suspecting a KL-based agreement measure would work better.
More from Research
- EMNLP 2026 industry track: keynotes on agentic challenges and constitutional alignment — verena_rieser · 2026-10-22
- Veteran reviewer: 'safe' cockroach papers get accepted, reviewers should reward flawed insight — plopesresearch · 2026-10-22
- Hyperparameter Tuning Is Itself a Machine-Learning Problem: Grid vs Random vs Bayesian — bravo_abad · 2026-10-22
- Michael Levin polls the public: how do physical, living, and minds overlap? — drmichaellevin · 2026-10-22
- ICLR reviewer: 2 of 3 assigned papers are complete AI slop, worst batch ever seen — AkariAsai · 2026-10-22
- Training-free 20B model Viveki generates real-time video heatmaps under 150ms — not-ekalabya · 2026-10-22