Teacher top-K distillation drops the low-probability <tool call> token from the response teacher's support, creating a one-sided gradient that causally raises tool over-calling.
Stabilizing On-Policy Distillation for MLLM Reasoning with Global Normalization
1 Pith paper cite this work. Polarity classification is still indexing.
abstract
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.
fields
cs.CL 1years
2026 1verdicts
ACCEPT 1representative citing papers
citing papers explorer
-
When Top-K Misses the Decision: Tool-Call Drift in Multi-Teacher On-Policy Distillation
Teacher top-K distillation drops the low-probability <tool call> token from the response teacher's support, creating a one-sided gradient that causally raises tool over-calling.