The paper introduces S$^3$T, a fully self‑contained framework for continuous video state tracking that uses temporal self‑distillation. It treats denser temporal sampling as privileged information, letting a dense‑view teacher guide a sparse‑view student to match its next‑token distribution without external labels or reward signals. Experiments on LLaVA-OneVision-2-8B show significant accuracy gains on VSTAT and MVBench benchmarks, and the learned capability transfers from synthetic to real videos.
We propose OPSD-V, an on-policy self-distillation paradigm for post-training few-step autoregressive (AR) video diffusion models. Existing few-step AR video generators can produce long videos with low latency, but still suffer from error accumulation and weakened motion dynamics during long autoregressive rollout.
Video-OPSD introduces a post‑training framework for Video Large Language Models that leverages privileged visual evidence to enhance on‑policy self‑distillation. The method constructs a self‑teacher conditioned only on annotated evidence frames, while the student processes the full video, allowing the teacher to provide more focused supervision. Additionally, an evidence‑guided token optimization scheme weights distillation based on each token’s reliance on privileged evidence, improving perceptually grounded reasoning. Experiments demonstrate consistent gains over standard OPSD and comparable performance to GRPO with less training time.
LeVJEPA is a video encoder that eliminates the need for architectural asymmetries, exponential-moving-average target encoders, stop-gradients, and capacity-limited predictors used in prior self‑supervised methods. It trains a single encoder with an invariance loss over global and local views, regularized by SIGReg to prevent collapse, and achieves strong performance with far less pretraining compute. The approach also allows block‑causal attention, making temporal ordering a property of the encoder itself, and matches or surpasses state‑of‑the‑art baselines on both appearance‑centric and motion‑centric benchmarks.
The paper introduces Privileged Appearance Transfer for Tracking (PATT), a teacher‑student framework that leverages exact target crops from past, current, and future frames during training to improve visual tracking. By weighting the teacher’s guidance with its localization advantage and accuracy, PATT transfers privileged appearance information to a deployable tracker that only uses past‑frame templates at inference. Experiments on seven benchmarks show consistent performance gains across both long‑ and short‑term tracking protocols.
arXiv:2609.09300v1 Announce Type: new Abstract: Video understanding demands a convergence of complementary capabilities across perception, temporal understanding, and complex reasoning, which are dif...
Vision-language models (VLMs) have made substantial progress in long-video understanding, with standard backbone models typically answering questions from frames sampled across the full video. However...
arXiv:2605. 23045v2 Announce Type: replace-cross Abstract: Video representation learning has seen tremendous progress in recent years.
The paper introduces Clue-OPSD, a clue‑privileged on‑policy self‑distillation framework that improves long‑video understanding by focusing on short, question‑relevant clue intervals rather than the entire video. Experiments on multiple benchmarks and Qwen3.5 model scales show that this approach consistently outperforms standard backbone models and competes strongly with supervised post‑training baselines, all while requiring fewer input frames and no additional inference modules.
arXiv:2608.22364v1 Announce Type: new Abstract: World action models (WAMs) couple visual future prediction with robot action generation, but accelerated students can lose task capabilities during dis...
arXiv:2609.05533v1 Announce Type: cross Abstract: Long-horizon manipulation is partially observable: the information needed to choose the next action may appear only in observations from minutes earl...
MAETrack introduces a lightweight framework to adapt pretrained masked autoencoder (MAE) representations for 3D single object tracking (SOT). It uses Layer‑Selective Initialization (LSI) to keep shallow geometric layers from the pre‑training while re‑initializing deeper layers, and Geometric Residual Gating (GRG) to emphasize salient regions in BEV features before template‑search fusion. Experiments on standard 3D SOT benchmarks show consistent improvements over vanilla fine‑tuning with minimal computational cost.
arXiv:2607. 23125v1 Announce Type: new Abstract: Post-training enables vision-language models (VLMs) to understand human instructions and perform various downstream tasks.