Knowledge Distillation (KD) aims to train a student model to mimic the behavior of a more powerful teacher model. In this paper, we reveal that through the lens of diffusion processes, student logits can be statistically treated as a noisy version of teacher logits, and KD helps reduce the noise level of student logits. This insight motivates us to design a framework leveraging KD to produce denoisable student logits that can be further recovered towards teacher logits via a reverse diffusion process. A key advantage of this approach is that the inference-diffusion process can occur in two physical locations and on separate devices, enabling a two-step and distributed inference process. The experimental results show that the derived denoisable student logits achieve comparable or even superior performance to standard KD’s, and the reverse diffusion process achieves a substantial improvement in accuracy, without needing the original image, thus preserving the privacy and security of the original data. Additionally, the logits can be further compressed before transmission, reducing the required bandwidth while achieving comparable overall performance.

Recent findings suggest that with the same teacher architecture, a fully converged or “stronger” checkpoint surprisingly leads to a worse student. This can be explained by the Information Bottleneck (IB) principle, as the features of a weaker teacher transfer more “dark” knowledge because they maintain higher mutual information with the inputs. Meanwhile, various works have shown that severe teacher-student structural disparity or capability mismatch often leads to worse student performance. To deal with these issues, we propose a generalizable and efficient Knowledge Distillation (KD) framework with implicit Student-Aware Teacher Ensembles (SATE). The SATE framework simultaneously trains a student network and a student-aware intermediate teacher as a learning companion. With the proposed co-training strategy, the intermediate teacher is trained gradually and forms implicit ensembles of weaker teachers along the learning process. Such a design enables the student model to retain more dark knowledge for better generalization ability. The proposed framework improves the training scheme in a plug-and-play way so that it can be applied to improve various classic and state-of-the-art KD methods on both intra-domain (up to 2.184 % ) and cross-domain (up to 7.358 % ) settings, under a diversified configurations on teacher-student architectures, and achieves a major efficient advantage over other generic frameworks. The code is available at https://github.com/diqichen91/SATE.git .

Despite recent advances, Text-to-video retrieval (TVR) is still hindered by multiple inherent uncertainties, such as ambiguous textual queries, indistinct text-video mappings, and low-quality video frames. Although interactive systems have emerged to address these challenges by refining user intent through clarifying questions, current methods typically rely on heuristic or ad-hoc strategies without explicitly quantifying these uncertainties, limiting their effectiveness. Motivated by this gap, we propose UMIVR, an Uncertainty-Minimizing Interactive Text-to-Video Retrieval framework that explicitly quantifies three critical uncertainties-text ambiguity, mapping uncertainty, and frame uncertainty-via principled, training-free metrics: semantic entropy-based Text Ambiguity Score (TAS), Jensen-Shannon divergence-based Mapping Uncertainty Score (MUS), and a Temporal Quality-based Frame Sampler (TQFS). By adaptively generating targeted clarifying questions guided by these uncertainty measures, UMIVR iteratively refines user queries, significantly reducing retrieval ambiguity. Extensive experiments on multiple benchmarks validate UMIVR's effectiveness, achieving notable gains in Recall@1 (69.2\% after 10 interactive rounds) on the MSR-VTT-1k dataset, thereby establishing an uncertainty-minimizing foundation for interactive TVR.

Conventional image restoration methods often implicitly assume that the degradation type in the input image is "seen" and "known" to the model, meaning it is trained and tested on the same type of degradation. More recent "all-in-one" models are designed to handle only one single degradation type in an image at a time, though the type can vary within a small, predefined set. This paper proposes OmniRestore, a novel approach to tackle a new and challenging task: "Omni Restoration", meaning restoring images with random, combined degradations of unspecified numbers and types. In this task, the restoration model must be able to restore images corrupted by multiple degradation types simultaneously, without prior knowledge of the exact types and the number of degradations in the input image. To address this, we devise a Mixture-of-Experts (MoE) architecture with a shared encoder and a group of type-sensitive decoder experts, alongside a two-stage training pipeline to expand the generalizability to various degradation types and their combinations. Extensive experiments demonstrate that our OmniRestore model consistently and significantly outperforms all state-of-the-art (SOTA) single-degradation models, vertical ensembles of those models, and "all-in-one" models on the Omni Restoration task. Our model also surpasses most of the competing models under a single-degradation setting with seen or unseen degradations. Our dataset and code are publicly available at https://github.com/anjusreekarnavar/OmniRestore.

We reveal that feedforward network (FFN) layers, rather than attention layers, are the primary contributors to Vision Transformer (ViT) inference latency, with their impact signifying as model size increases. This finding highlights a critical opportunity for optimizing the efficiency of large-scale ViTs by focusing on FFN layers. In this work, we propose a novel channel idle mechanism that facilitates post-training structural reparameterization for efficient FFN layers during testing. Specifically, a set of feature channels remains idle and bypasses the nonlinear activation function in each FFN layer, thereby forming a linear pathway that enables structural reparameterization during inference. This mechanism results in a family of **RePa**rameterizable **Vi**sion **T**ransformers (RePaViTs), which achieve remarkable latency reductions with acceptable sacrifices (sometimes gains) in accuracy across various ViTs. The benefits of our method scale consistently with model sizes, demonstrating greater speed improvements and progressively narrowing accuracy gaps or even higher accuracies on larger models. In particular, RePa-ViT-Large and RePa-ViT-Huge enjoy **66.8%** and **68.7%** speed-ups with **+1.7%** and **+1.1%** higher top-1 accuracies under the same training strategy, respectively. RePaViT is the first to employ structural reparameterization on FFN layers to expedite ViTs to our best knowledge, and we believe that it represents an auspicious direction for efficient ViTs. Source code is available at https://github.com/Ackesnal/RePaViT.

Generalist robot manipulation policies (GMPs) have the potential to generalize across a wide range of tasks, devices, and environments. However, existing policies continue to struggle with out-of-distribution scenarios due to the inherent difficulty of collecting sufficient action data to cover extensively diverse domains. While fine-tuning offers a practical way to quickly adapt a GMPs to novel domains and tasks with limited samples, we observe that the performance of the resulting GMPs differs significantly with respect to the design choices of fine-tuning strategies. In this work, we first conduct an indepth empirical study to investigate the effect of key factors in GMPs fine-tuning strategies, covering the action space, policy head, supervision signal and the choice of tunable parameters, where 2,500 rollouts are evaluated for a single configuration. We systematically discuss and summarize our findings and identify the key design choices, which we believe give a practical guideline for GMPs fine-tuning. We observe that in a lowdata regime, with carefully chosen fine-tuning strategies, a GMPs significantly outperforms the state-of-the-art imitation learning algorithms. The results presented in this work establish a new baseline for future studies on fine-tuned GMPs.

Despite the significant success of open-vocabulary large foundation models (LFM) for segmentation in recent years, most real-life applications remain closed-vocabulary tasks and efficiency remains a critical factor for usability. Leveraging the power of open-vocabulary LFMs to create efficient, accurate closed-vocabulary segmentation models without the burden of pixel annotations provides an essential tool for employing these models effectively. This work introduces a novel Open-vocabulary to Closed-vocabulary Segmentation (O2CSeg) framework, which builds a compact, closed-vocabulary segmentation model from a large open-vocabulary LFM without any annotations. Our method capitalises on AI-assisted text prompts and learnable prompts that correspond to target class names to fully unleash the potential of open-vocabulary LFMs, eliminating the expensive annotation process. To address the challenge of noisy pseudo-labels generated by the teacher during training, we propose a confidence margin-based re-weighting knowledge distillation scheme, ensuring the student captures high-quality knowledge from the teacher. We validate our framework across diverse datasets and student configurations, demonstrating its efficacy in achieving high efficiency and accuracy for segmentation without annotations. In many cases, the derived student network outperforms its open-vocabulary teacher significantly with a higher mIoU and 40 times faster speed. Our contributions offer a promising direction for fast prototyping of efficient semantic segmentation models in scenarios where annotations are lacking, or label sets are evolving.

We present Chain-of-Action (CoA), a novel visuomotor policy paradigm built upon Trajectory Autoregressive Modeling. Unlike conventional approaches that predict next step action(s) forward, CoA generates an entire trajectory by explicit backward reasoning with task-specific goals through an action-level Chain-of-Thought (CoT) process. This process is unified within a single autoregressive structure: (1) the first token corresponds to a stable keyframe action that encodes the task-specific goals; and (2) subsequent action tokens are generated autoregressively, conditioned on the initial keyframe and previously predicted actions. This backward action reasoning enforces a global-to-local structure, allowing each local action to be tightly constrained by the final goal. To further realize the action reasoning structure, CoA incorporates four complementary designs: continuous action token representation; dynamic stopping for variable-length trajectory generation; reverse temporal ensemble; and multi-token prediction to balance action chunk modeling with global structure. As a result, CoA gives strong spatial generalization capabilities while preserving the flexibility and simplicity of a visuomotor policy. Empirically, we observe that CoA outperforms representative imitation learning algorithms such as ACT and Diffusion Policy across 60 RLBench tasks and 8 real-world tasks.

Collecting and curating data plays a crucial role in environmental surveying. In order to gather meaningful samples, it is often necessary to develop a real-time data curation system that processes data on-the-fly and enriches it with validated models from domain experts. One area where this is particularly important is underwater marine surveys, where the vastness of the sea requires human interaction in the curation process to focus on relevant areas for exploration. Additionally, ongoing surveys are susceptible to poor performance due to data drift, which hinders the ability to provide valuable feedback for guiding data collection. While recent advancements have shown promise in addressing these challenges, they often overlook the practical constraints associated with remote data collection, such as limited processing power and latency. To overcome these limitations, this paper proposes a real-time system that adapts to data drift and enables the recording of uncertain samples for further processing and analysis on shore. The results of our approach demonstrate a remarkable improvement in species recognition, achieving an almost 18% improvement compared to the best baseline method in unseen areas. Importantly, this improvement is achieved while meeting the real-time requirements of surveys and consuming only 15W of power. By effectively addressing the challenges of underwater data drift, our proposed approach provides an efficient and effective solution for environmental surveys.

Machine learning (ML) algorithms have shown great potential in edge-computing environments, however, the literature mainly focuses on model inference only. We investigate how ML can be operationalized and how in-situ curation can improve the quality of edge applications, in the context of ML-assisted environmental surveys. We show that camera-enabled ML systems deployed on edge devices can enable scientists to perform real-time monitoring of species of interest or characterization of natural habitats. However, the benefit of this new technology is only as good as the quality and accuracy of the edge ML model inferences. In this demonstration, we show that with small additional time investment, domain scientists can manually curate ML model outputs and thus obtain highly reliable scientific insights, leading to more effective and scalable environmental surveys.

Crown-of-Thorn Starfish (COTS) outbreaks are a major cause of coral loss on the Great Barrier Reef (GBR) and substantial surveillance and control programs are ongoing to manage COTS populations to ecologically sustainable levels. In this paper, we present a comprehensive real-time machine learning-based underwater data collection and curation system on edge devices for COTS monitoring. In particular, we leverage the power of deep learning-based object detection techniques, and propose a resource-efficient COTS detector that performs detection inferences on the edge device to assist marine experts with COTS identification during the data collection phase. The preliminary results show that several strategies for improving computational efficiency (e.g., batch-wise processing, frame skipping, model input size) can be combined to run the proposed detection model on edge hardware with low resource consumption and low information loss.