DOPD: Dual On-policy Distillation
arXiv · HuggingFace · ▲93
Abstract (verbatim)
On-policy distillation (OPD) offers superior capacity transfer by supervising student-sampled trajectories with dense token-level signals. To furnish high-quality supervision sources and thereby elevate the performance frontier of distillation, an intuitive direction is to infuse privileged information to either teacher or student itself. However, this additional input induces a potential failure mode we dub privilege illusion: a pattern that conflates the transferable capability gap that students are meant to close, and the information asymmetry gap that can only be mimicked but never replicated. This issue is further amplified by the inherent non-uniformity of token-level supervision, where only a small subset of tokens carries pivotal capability-bearing signals. To this end, we propose DOPD, an advantage-aware dual distillation paradigm that dynamically routes token-level supervision between privileged teacher and privileged student policies based on their advantage gap and relative probabilities. Each token receives supervision of different strength, objective, and strategy from either teacher or student itself, which transfers credible capability while simultaneously receiving auxiliary signals, to alleviate privilege illusion. Extensive experiments on both large language model (LLM) and vision-language model (VLM) settings demonstrate that DOPD consistently outperforms Vanilla OPD and other counterparts. Further results on stability, robustness, continual learning, and out-of-distribution tasks validate its superiority.
Background
To understand this research, we first clarify the technical background: Model distillation, a core technique for transferring capabilities from a "teacher" to a "student" model, is widely used in optimizing large language models (LLMs) and vision-language models (VLMs). Traditional methods rely on "off-policy trajectories," but student models' behavior distributions may diverge from teachers. The emerging "on-policy distillation" (OPD) addresses this by having student models generate their own trajectories and using teachers' token-level supervision, improving efficiency and performance. However, OPD's upper performance limit is constrained by supervision quality—when "privileged information" (e.g., reasoning hints for LLMs or structured annotations for VLMs) is introduced, it may create "privilege illusion": the student mimics surface advantages from information asymmetry rather than truly transferable abilities.
Previous limitations stem from uniform supervision strategies. Existing OPD methods treat all tokens equally, ignoring their value differences. For example, some tokens carry critical capability signals (e.g., logical reasoning steps), while others are noisy. When privileged information is added, this uniform supervision causes students to overfit "privilege shortcuts" instead of core abilities, leading to performance instability, reduced exploration, and training collapse.
This paper's solution is "Dual On-policy Distillation" (DOPD). Its core idea is to dynamically differentiate token supervision sources: For tokens where the teacher has a credible capability advantage, stronger teacher supervision is used to transfer high-value signals; for tokens dominated by privileged information or irrelevant to core abilities, lighter supervision (or student self-supervision) is applied to avoid overfitting. DOPD adjusts supervision strength, objectives, and strategies for each token based on "advantage gaps" (between teacher and student) and "predicted probabilities," reducing privilege illusion.
Key differences from prior work: Unlike Vanilla OPD, DOPD does not assume all tokens contribute equally to capability transfer. It selectively allocates supervision resources, combining teacher supervision (for transferable privileges) and student self-supervision (for stability/exploration) via dynamic routing. This not only improves performance (significantly outperforming Vanilla OPD on LLM/VLM tasks) but also enhances robustness, continual learning, and out-of-distribution generalization. Its innovation lies in shifting privileged information use from "blind imitation" to "selective absorption," addressing OPD's core flaws.




