Student driven curriculum cuts training data and speeds video grounding
Less Data, Better Timing: Student-Curriculum Coupling for VLM On-Policy Distillation in Temporal Video Grounding
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Temporal video grounding means teaching computers to find when things happen in videos based on language. Training these systems usually needs lots of examples from a fixed teacher, but not all examples help at every learning step. The authors show that matching examples to what the student model currently understands makes training more efficient. Their new approach lets the student pick which training examples to focus on, leading to faster learning using fewer data while getting better at finding events in videos.
What this means in practice
- •For video ai developers: Train video understanding models faster with fewer labeled examples by using a student-driven curriculum approach.
- •For machine learning engineers: Reduce computation time and data needs when refining vision-language models for temporal event localization by dynamically adjusting training samples.
Authors
Jiacheng Qiu, Yunsoo Kim, Ruichen Xu, Jian Luo, Petar M. Djurić, Sima Mofakham
Abstract
On-policy distillation (OPD) provides dense supervision directly on student-generated trajectories, making it an effective post-training strategy for vision-language models in temporal video grounding (TVG). However, existing pipelines typically construct the training curriculum from a fixed teacher and the initial student state, implicitly assuming that selected examples retain positive supervision value throughout optimization. We show that supervision trustworthiness and supervision necessity are distinct yet coupled: the former concerns target credibility, while the latter varies with the student's current task competence; together, they shape supervision value. Building on this coupled view, we introduce Student-Curriculum Coupling (SCC), a closed-loop framework in which a compact Anchor-Frontier curriculum defines the candidate supervision space and the evolving student dynamically determines its active subset. Supervision can therefore be activated, suspended, or reactivated as competence changes, concentrating teacher computation and optimization on current task-level deficits. Across three TVG benchmarks, SCC achieves a 5.1% relative improvement in mean recall over Video-OPD on its original curriculum, while using 60.0% fewer training examples and reducing training time by 50.4%. Ablations support the complementary roles of capability-structured curriculum design and student-dependent supervision in achieving these gains. Together, these results establish SCC as a data- and compute-efficient framework for TVG post-training, delivering stronger temporal grounding by aligning trustworthy supervision with the student's evolving learning needs.