Share your thoughts, 1 month free Claude Pro on usSee more
WorkDL logo mark

Stabilizing On-Policy Distillation for MLLM Reasoning with Global Normalization

About

On-policy distillation (OPD) has recently emerged as an important post-training paradigm. By using a stronger teacher model to provide dense, fine-grained supervision for sampled trajectories, OPD offers a clear advantage over reinforcement learning with verifiable rewards (RLVR), which typically depends on sparse binary or outcome-based environmental feedback. However, naive token-level distillation can suffer from gradient instability, due to magnitude misalignment in outlier states. To address this issue, we propose Globally Normalized Distillation Policy Optimization (GNDPO), a practical method that stabilizes optimization by transforming raw KL scores into batch-level relative advantages. This normalization effectively mitigates gradient explosions while retaining the benefits of token-level guidance. Experimental results show that GNDPO substantially improves training robustness and downstream performance across multimodal reasoning tasks. The code is released at https://github.com/OPPO-Mente-Lab/GNDPO.

Dongze Hao, Zhiwei Jin, Chen Chen, Haonan Lu• 2026

Related benchmarks

TaskDatasetResultRank
Multimodal Math ReasoningMathVision
Accuracy51.6
263
Multimodal Math ReasoningWeMath
Accuracy48.8
228
Multimodal Mathematical ReasoningMathVista mini
Accuracy0.73
124
Multimodal Logical ReasoningLogicVista
Accuracy54.4
76
Massive Multi-discipline Multimodal UnderstandingMMMU (val)--
18
Multimodal Mathematical ReasoningMathVerse Vision Only
Accuracy53.2
13
Multimodal Mathematical ReasoningDynaMath worst case
Accuracy32.5
13
Showing 7 of 7 rows

Other info

Follow for update