Ilmu Komputer & AI editorial
Open AccessOA2026
BAS-OPD: Budget-Aware Selective On-Policy Self-Distillation for Fine-Grained Multimodal Perception
Selective teacher supervision under limited query budgets for fine-grained multimodal perception
Zihan Chen; Hengguang Zhou; Yuan Kang; Yiming Zhang; Wenhui Fang; Zenghui Ding; Yining Sun; Cho-Jui Hsieh· 2026· DOI 10.48550/arXiv.2609.25891
The core problem
Multimodal large language models (MLLMs) often struggle with fine-grained visual perception when processing complete images, because critical evidence may only appear in local regions. On-policy self-distillation (OPD) enables transferring privileged visual knowledge from informative views to full-image policies, but querying the teacher for every rollout introduces substantial supervision costs. This creates a tension between the benefits of privileged-view supervision and the computational expense of teacher queries. The paper proposes BAS-OPD, a budget-aware selective OPD framework that allocates teacher supervision under limited query budgets. Instead of querying all rollouts, BAS-OPD selects informative samples while maintaining full-batch student generation. The authors explore random, uncertainty-based, and learned utility-based selection strategies. The learned selector estimates query value from detached rollout statistics and online utility signals derived from student–teacher agreement and teacher confidence without additional student forward passes. BAS-OPD only changes training-time supervision allocation and preserves single-pass full-image inference. Experiments on f
Innovation
The authors evaluate BAS-OPD on fine-grained multimodal perception benchmarks. While the abstract does not provide specific numbers, it states that BAS-OPD achieves strong performance while substantially reducing teacher supervision costs. The experiments compare the three selection strategies (random, uncertainty-based, learned utility-based) under varying query budgets. The results demonstrate that selective OPD under constrained budgets is effective, and that the learned utility-based selection likely outperforms the baselines by focusing teacher supervision on the most informative rollouts. The paper highlights that BAS-OPD preserves single-pass full-image inference, meaning that at test time, the model does not require multiple views or additional computation. This is crucial for practical deployment. The reduction in teacher supervision costs is substantial, as only a fraction of rollouts are queried. The exact budget levels and performance metrics are detailed in the full paper, but the abstract confirms that the approach yields strong performance across benchmarks.
Multimodal large language models (MLLMs) often struggle with fine-grained visual perception when processing complete images, because critical evidence may only appear in local regions. On-policy self-distillation (OPD) enables transferring privileged visual knowledge from informative views to full-image policies, but querying the teacher for every rollout introduces substantial supervision costs. This creates a tension between the benefits of privileged-view supervision and the computational expense of teacher queries. The paper proposes BAS-OPD, a budget-aware selective OPD framework that allocates teacher supervision under limited query budgets. Instead of querying all rollouts, BAS-OPD selects informative samples while maintaining full-batch student generation. The authors explore random, uncertainty-based, and learned utility-based selection strategies. The learned selector estimates query value from detached rollout statistics and online utility signals derived from student–teacher agreement and teacher confidence without additional student forward passes. BAS-OPD only changes training-time supervision allocation and preserves single-pass full-image inference. Experiments on fine-grained multimodal perception benchmarks demonstrate that BAS-OPD achieves strong performance while substantially reducing teacher supervision costs, highlighting the effectiveness of selective OPD under constrained budgets.
BAS-OPD operates by decoupling student generation from teacher supervision. The student model generates rollouts for the full batch, but only a selected subset of these rollouts is queried to the teacher for supervision. This selective allocation is governed by a budget constraint: the number of teacher queries is limited to a fraction of the full batch. The framework considers three selection strategies:
Why it matters
The key insight of BAS-OPD is that not all rollouts are equally informative for distillation. By selectively querying the teacher, the method reduces computational cost while maintaining or improving performance. The learned utility-based selection is particularly promising because it adapts to the student's current state and the teacher's reliability. The use of detached rollout statistics and online utility signals avoids extra forward passes, making the selection overhead minimal. The framework is general and could be applied to other on-policy distillation settings. However, the effectiveness of the learned selector depends on the quality of the utility signals, and there may be a trade-off between the budget and the performance gain. Future work could explore more sophisticated selection strategies, adaptive budgeting, and extension to other modalities. The paper's contribution lies in formalizing the budget-aware selective OPD problem and providing a practical solution that maintains single-pass inference. This aligns with the growing need for efficient multimodal models that can handle fine-grained perception without excessive computational overhead.
Who should read this
CS practitioners and researchers
Opening member content…