Search papers, labs, and topics across Lattice.
This study systematically investigates the dynamics of on-policy distillation (OPD) in large language models, identifying its role as an exploration catalyst that guides student models toward correct reasoning paths. The authors reveal that the effectiveness of OPD is critically dependent on the quality of the guiding signal, exposing two significant pathologies: the Student-Teacher Mismatch and Length Exploitation, which can derail effective exploration. By implementing lightweight signal regulations such as advantage clipping and log-scale compression, the researchers demonstrate improved exploration and distillation performance across multiple benchmarks, highlighting the importance of signal quality over teacher model size.
The effectiveness of on-policy distillation hinges not on the size of the teacher model, but on the quality of the guiding signal, revealing critical pathologies that can undermine exploration.
On-policy distillation (OPD) has become a key paradigm in LLM post-training, yet its training dynamics remain poorly understood. We present a systematic study examining the role, pathologies, and regulations of OPD. We first clarify the role of OPD as an exploration catalyst: it steers the student toward correct reasoning paths via dense token-level guidance, without expanding capability ceiling. We confirm this by showing that prompt diversity matters more than per-problem sampling numbers, and critically, that the effectiveness of OPD hinges entirely on the quality of its guiding signal. This dependency exposes two pathologies that derail exploration. The Student-Teacher Mismatch occurs when a large teacher-student distributional gap causes the guiding signal to misalign with task correctness, steering exploration in counterproductive directions. Length Exploitation arises when the aggregated token-level objective creates length-dependent shortcuts, allowing the student to game the reward landscape through response truncation or redundant padding, exploring degenerate length modes rather than reasoning strategies. To tame these pathologies, we investigate lightweight signal regulations: advantage clipping and log-scale compression, ensuring exploration is guided by faithful signals. Experiments across seven benchmarks demonstrate that these regulations alleviate length exploitation and enable effective distillation, stably surpassing OPD variants and RLVR baselines, thereby confirming that well-regulated signal quality, rather than mere teacher scale, governs successful exploration in OPD.
