Recent advancements in multimodal large language models (MLLMs) and video agent systems have significantly improved general video understanding. However, when applied to scientific video understanding and educating—a domain that demands external professional knowledge integration and rigorous step-wise reasoning—existing approaches often struggle. To bridge this gap, we propose SciEducator, an iterative self-evolving multi-agent system for scientific video comprehension and education. Rooted in the classical Deming Cycle from management science, our design reformulates its Plan–Do–Study–Act philosophy into a self-evolving reasoning and feedback mechanism, which facilitates the interpretation of intricate scientific activities in videos. Moreover, SciEducator can produce multimodal educational content tailored to specific scientific processes, including textual instructions, visual guides, audio narrations, and interactive references. To support evaluation, we construct SciVBench, a benchmark consisting of 500 expert-verified and literature-grounded science QA pairs across five categories, covering physical, chemical, and everyday phenomena. Extensive experiments demonstrate that SciEducator substantially outperforms leading closed-source MLLMs (e.g., Gemini, GPT-4o) and state-of-the-art video agents on the benchmark, establishing a new paradigm for the community.

Reliable visual safety understanding in real-world scenarios demands more than just object recognition; it requires causal reasoning under epistemic uncertainty. While Large Vision-Language Models (LVLMs) demonstrate impressive semantic alignment on standard benchmarks, they often struggle to distinguish between superficial correlation and genuine forensic logic when grounded in the dynamic, partially observable nature of first-person experiences. Existing evaluations, dominated by third-person surveillance footage and binary classification metrics, fail to expose this cognitive gap. To address this, we introduce EgoSafe-Bench, a benchmark specifically designed to probe forensic reasoning in egocentric safety scenarios. It comprises 12,000 unique evaluation samples, generated by pairing each of the 3,000 video clips with a QA chain governed by our proposed Hierarchical Reasoning Evaluation (HRE) protocol. Unlike standard benchmarks, HRE mandates a rigorous reasoning trajectory from initial feature anchoring to blind-spot deduction and intent inference, thereby enforcing logical consistency and penalizing shortcut-based predictions.Extensive evaluations of state-of-the-art LVLMs (e.g., Qwen3-VL, Gemini, VideoLLaMA 3) reveal a significant perception-reasoning decoupling: models often achieve high descriptive scores but exhibit notable fragility in causal reasoning and logical closure. Our work provides both a challenging dataset and a systematic evaluation framework to foster the development of logically robust video understanding systems.

Video transformers (VidTs) have demonstrated powerful spatio-temporal modelling capabilities in various video understanding tasks. However, their practical deployment remains constrained by high computational and memory overheads, primarily caused by the quadratic complexity of spatio-temporal self-attention and the intensive demands of processing high-frame-rate video streams. While recent efforts have attempted to leverage spatio-temporal similarity to eliminate redundant computation, they either rely on expensive similarity recognition or support only limited reuse, resulting in limited performance gains. In this work, we present a comprehensive analysis of the potential opportunities and challenges in exploiting spatiotemporal similarity within VidTs. Building upon these insights, we propose SimVidT, a novel VidT accelerator that leverages spatio-temporal similarity to reduce computational redundancy. SimVidT is supported by an algorithm and architecture codesign. At the algorithm level, we introduce a locality-aware similarity identification method and a redundancy-eliminated attention mechanism to efficiently identify and eliminate redundant computations at multiple granularities. At the architecture level, we develop a dedicated accelerator featuring a customized hierarchical memory management unit and specialized compute units, which adapts to the complex execution flow of VidTs and translates redundancy reduction into performance gains. We implement a cycle-accurate simulator for performance evaluation. Experimental results demonstrate that SimVidT achieves 1.7×speedup and 2.5× better energy efficiency on average compared to the state-of-the-art VidT accelerator CMC.

State space models (SSMs) have emerged as a powerful paradigm for efficient single-image super-resolution (SR) due to their linear complexity and long-range modeling capabilities. However, existing Mamba-based methods typically rely on data-agnostic rigid scanning, which reshapes 2D images into 1D sequences over a fixed grid, inevitably disrupting spatial-semantic topology and introducing artifacts. Inspired by the Gestalt perceptual grouping theory, we propose SP-MoMamba, a superpixel-driven mixture of state space experts designed for content-aware SR. Our core idea is to transform the traditional rigid scanning into a semantic-level interaction by treating superpixels as fundamental units. Specifically, we introduce the Superpixel-driven State Space Model (SP-SSM), which compresses semantically homogeneous regions into high-order tokens to preserve global topological consistency. To address the conflict between fixed scanning scales and diverse semantic granularities, we develop the Multi-Scale Superpixel Mixture of State Space Experts (MSS-MoE). This module utilizes a dynamic routing mechanism to adaptively assign scale-specific experts, effectively capturing multi-scale textures while reducing computational redundancy. Furthermore, to prevent the loss of high-frequency details during global abstraction, we introduce a Local Spatial Modulation Expert (LSME) to complement the global modeling, ensuring a precise reconstruction of sharp edges and fine structures. Extensive experiments on standard benchmarks demonstrate that SP-MoMamba achieves superior reconstruction fidelity and a more favorable efficiency-performance trade-off compared to state-of-the-art efficient SR methods.

With the growing complexity of Large Language Models (LLMs), there is increasing interest in efficient inference architectures. Early exiting is a popular method that improves inference efficiency by skipping layers and generating output once the model reaches a certain confidence level. Traditional early exiting methods incorporate weighted cross-entropy loss during training, ensuring accurate predictions across all internal classifiers. However, only one correct prediction is needed to speed up the processing during inference. Additionally, early exiting in LLMs also faces challenges with KV cache updates in autoregressive decoding. Current solutions approximate the KV cache of skipped layers by copying hidden states, resulting in reduced accuracy. In response, we propose ConsistentEE, an enhanced early exiting framework that maintains consistency between training and inference. Our method frames early exiting as a reinforcement learning challenge, employing policy networks to ascertain when to exit. We also introduce Memory Layer to evaluate the hardness of instances. To tackle KV cache challenges in LLMs with early exiting, we employ a parallel decoding strategy. Experimental results show that our approach consistently outperforms the baseline approach on multiple natural language classification and generation tasks.

Human-object interaction (HOI) detection is crucial for advanced scene understanding, but real-world deployment requires models that continually acquire new interactions without forgetting previous knowledge. This incremental setting is especially challenging in HOI due to two coupled issues: catastrophic forgetting and severe multi-label long-tailed imbalance, further exacerbated in privacy-sensitive applications where revisiting past data is infeasible. In this paper, we introduce Prototype-Guided Routed Analytic Learning (PGRAL), an exemplar-free framework that jointly addresses both challenges. PGRAL first establishes a highly discriminative representation space via prototype-guided contrastive refinement for pre-training, enhancing intra-class compactness and inter-class separability. For incremental learning, our routed analytic incremental learner recursively updates the classifier in closed-form, effectively eliminating task-recency bias and catastrophic forgetting. To address the long-tailed imbalance, we further devise a knowledge-enhanced feature synthesis module that generates high-quality features for tail classes, with the learner’s routing mechanism ensuring safe and effective integration. Extensive experiments on HICO-DET and V-COCO demonstrate that PGRAL achieves a new state-of-the-art performance, with particularly significant gains on rare tail interactions, offering a powerful and scalable solution for HOI understanding in dynamic, real-world environments. Codes will be publicly available.

Low-rank adaptation (LoRA) is a widely used parameter-efficient fine-tuning method that places trainable low-rank adapters into frozen pre-trained models. Recent studies show that using fewer LoRA adapters may still maintain or even improve performance, but existing methods still distribute adapters broadly, leaving where to place a limited number of adapters to maximize performance largely open. To investigate this, we introduce PAGE (Projected Adapter Gradient Energy), a gradient-based sensitivity probe that estimates the initial trainable gradient energy available to each candidate LoRA adapter. Surprisingly, we find that PAGE is highly concentrated on a single shallow FFN down-projection across two model families and four downstream tasks. We term this module the dominant adaptation module and show that its layer index is architecture-dependent but task-stable. Motivated by this finding, we propose DomLoRA, a placement method that places a single adapter at the dominant adaptation module. With only ~0.7% of vanilla LoRA's trainable parameters, DomLoRA outperforms it on average across various downstream tasks, including instruction following, mathematical reasoning, code generation, and multi-turn conversation. This method also improves other LoRA variants, supporting the dominant adaptation module perspective as a practical placement guideline.

Multiple-object tracking (MOT) involves identifying and consistently assigning IDs to multiple targets in a video sequence, where occlusions are frequently encountered. Recent approaches leverage appearance features and incorporate temporal information to improve discriminative power for better occlusion handling. However, most existing methods fail to fully utilize historical tracking information without significantly compromising tracking efficiency. To address this limitation, we propose a new MOT framework called the feature-adaptive continual learning tracker (FACT), which supports online learning while effectively leveraging the complete tracking history. At the core of this framework is the feature-adaptive continual-learning (FAC) module, which adopts an analytic continual learning to enable online training using complete historical tracking information. Specifically, FAC module updates appearance models using only current-frame features and a fixed-size feature auto-correlation matrix via closed-form update rules, achieving performance equal to joint training on all past tracking data. We show that the FAC module can be seamlessly integrated into existing feature-based trackers, enhancing performance with minimal impact on tracking speed. Additionally, we introduce a two-stage association mechanism tailored to the proposed analytic continual learning-based paradigm. Extensive experiment results demonstrate that the proposed method achieves state-of-the-art online tracking performance on MOT16, MOT17, and MOT20 benchmarks.

Exemplar-free class-incremental learning (EFCIL) aims to mitigate catastrophic forgetting in class-incremental learning (CIL) without relying on historical training samples as exemplars. Compared with exemplar-based CIL that stores exemplars, EFCIL is more prone to forgetting. An emergent EFCIL branch named Analytic Continual Learning (ACL) introduces a gradient-free paradigm based on Recursive Least-Square for forgetting-resistant classifier training with a frozen backbone during CIL. However, ACL currently suffers from ineffective representations and limited utilization of backbone knowledge. To address these challenges, we propose a representation-enhanced analytic learning (REAL) scheme. REAL improves the representation by constructing a dual-stream base pretraining stage followed by a representation-enhancing distillation process. The dual-stream base pretraining combines self-supervised contrastive learning for general features with supervised learning for class-specific knowledge, followed by representation-enhancing distillation to integrate both streams, improving representations for the subsequent CIL paradigm. REAL further introduces a feature fusion buffer to multi-layer backbone features, enabling richer feature extraction for classifier training. Our proposed method can be incorporated into existing ACL techniques, yielding superior performance. Empirical results demonstrate that REAL achieves state-of-the-art performance on CIFAR-100, ImageNet-100, and ImageNet-1k benchmarks, outperforming existing exemplar-free methods and rivaling exemplar-based approaches.

Long Chain-of-Thought (CoT) reasoning has significantly advanced the capabilities of Large Language Models (LLMs), but this progress is accompanied by substantial memory and latency overhead from the extensive Key-Value (KV) cache. Although KV cache quantization is a promising compression technique, existing low-bit quantization methods often exhibit severe performance degradation on complex reasoning tasks. Fixed-precision quantization struggles to handle outlier channels in the key cache, while current mixed-precision strategies fail to accurately identify components requiring high-precision representation. We find that an effective low-bit KV cache quantization strategy must consider two factors: a key channel's intrinsic quantization difficulty and its relevance to the query. Based on this insight, we propose MixKVQ, a novel plug-and-play method that introduces a lightweight, query-aware algorithm to identify and preserve critical key channels that need higher precision, while applying per-token quantization for value cache. Experiments on complex reasoning datasets demonstrate that our approach significantly outperforms existing low-bit methods, achieving performance comparable to a full-precision baseline at a substantially reduced memory footprint.

Multimodal Large Language Models (MLLMs) are increasingly vulnerable to multimodal Indirect Prompt Injection (IPI) attacks, which embed malicious instructions in images, videos, or audio to hijack model behavior. Existing defenses, designed primarily for text-only LLMs, are unsuitable for countering these multimodal threats, as they are easily bypassed, modality-dependent, or generalize poorly. Inspired by activation steering researches, we hypothesize that a robust, general defense independent of modality can be achieved by steering the model's behavior in the representation space. Through extensive experiments, we discover that the instruction-following behavior of MLLMs is encoded in a subspace. Steering along directions within this subspace can enforce adherence to user instructions, forming the basis of a defense. However, we also found that a naive defense direction could be coupled with a utility-degrading direction, and excessive intervention strength harms model performance. To address this, we propose ARGUS, which searches for an optimal defense direction within the safety subspace that decouples from the utility degradation direction, further combining adaptive strength steering to achieve a better safety-utility trade-off. ARGUS also introduces lightweight injection detection stage to to activate the defense on-demand, and a post-filtering stage to verify defense success. Experimental results show that ARGUS can achieve robust defense against multimodal IPI while maximally preserving the MLLM's utility.

In real-world applications, video action recognition models must continuously learn new action categories while retaining previously acquired knowledge. However, most existing approaches rely on storing historical data for replay, which introduces storage burdens and raises data privacy concerns. To address these challenges, we investigate the problem of Exemplar-Free Continual Video Action Recognition (EF-CVAR) and propose a novel framework named Slow-Fast Collaborative Learning (SFCL). SFCL integrates two complementary learning paradigms: a slow branch based on gradient-driven deep learning, which provides strong adaptability to new tasks, and a fast branch based on analytic learning (e.g., Recursive Least Squares), which efficiently preserves old knowledge without requiring access to past samples. To enable effective collaboration between the two branches, we design the Slow-Fast Dynamic Re-parameterization (SFDR) mechanism for adaptive fusion, and the Knowledge Reflection Mechanism (KRM), which mitigates forgetting and task-recency bias via pseudo-feature generation and dual-level knowledge distillation. Extensive experiments on UCF101, HMDB51, and Something-Something V2 demonstrate that SFCL achieves superior performance compared to existing replay-based methods, despite being exemplar-free. Notably, in long-duration continual learning scenarios, SFCL exhibits remarkable robustness, achieving up to a 30.39\% improvement in accuracy over baselines while maintaining a low forgetting rate, highlighting its scalability and effectiveness in real-world video recognition tasks.

Class-Incremental Learning (CIL) with pre-trained models (PTMs) aims to sequentially adapt PTMs to new categories without forgetting old knowledge. Built upon PTMs, existing adapter-based methods mainly train models via distinct task-specific adapters, and present a uniform knowledge allocation for each adapter during inference. However, this allocation mechanism ignores the nature of task discrepancy and leads to suboptimal utilization of adapters. Also, under CIL constraint, an allocator is prone to forgetting when tasks evolve. To address these issues, we propose a Non-Forgetting Allocation with Bi-Level Competition (NoFA-BC). NoFA-BC constructs a non-forgetting allocator (NFA) by transforming the allocator training into a recursive least-squares problem and achieves an allocator equivalent to that trained with all data. Based on the NFA, a Bi-Level Competition (BLC) including an intra-task level Winner-Takes-All (WTA) mechanism and inter-task Last-Ones-Fall (LOF) elimination is proposed to provide better allocation of adapter knowledge. WTA extracts the most significant logit within a task to represent the adapter's contribution and LOF suppresses the irrelevant adapters. With BLC, participation ratio of each adapter can be tailored for each input. Moreover, a Stability Enhancement (SE) process is incorporated to further improve the performance of old tasks.

Federated Learning (FL) is a popular distributed learning paradigm to break down data silo. Traditional FL approaches largely rely on gradient-based updates, facing significant issues about heterogeneity, scalability, convergence, and overhead, etc. Recently, some analytic-learning-based work has attempted to handle these issues by eliminating gradient-based updates via analytical (i.e., closed-form) solutions. Despite achieving superior invariance to data heterogeneity, these approaches are fundamentally limited by their single-layer linear model with a frozen pre-trained backbone. As a result, they can only achieve suboptimal performance due to their lack of representation learning capabilities. In this paper, to enable representable analytic models while preserving the ideal invariance to data heterogeneity for FL, we propose our Deep Analytic Federated Learning approach, named DeepAFL. Drawing inspiration from the great success of ResNet in gradient-based learning, we design gradient-free residual blocks in our DeepAFL with analytical solutions. We further introduce an efficient layer-wise protocol for training our deep analytic models layer by layer in FL through least squares. Both theoretical analyses and empirical evaluations validate our DeepAFL's superior performance with its dual advantages in heterogeneity invariance and representation learning, outperforming state-of-the-art baselines by up to 5.68%-8.42% across three benchmark datasets. The related codes will be made open-sourced upon the acceptance of this paper.

Unsupervised cross-modal hashing enables efficient retrieval of semantically related instances across different modalities without requiring manual semantic annotation. However, existing unsupervised methods rely heavily on large-scale image-text pairs. Collecting such data can be costly, particularly in scenarios where well-aligned pairs are scarce due to privacy and specialized constraints. More critically, existing methods tend to overfit to seen training data, restricting their generalization performance on unseen categories that the constrained training data cannot cover. To address these limitations, we propose Attribute-Prompted Kernel Hashing (APKH), a novel data-efficient approach that constructs a compact, modality-aligned Hamming space driven by the generalized attribute priors of vision-language foundation models. Specifically, APKH introduces two core modules: Context-optimized Attribute Kernel Mapping (CAKM) and Kernel-Smoothed Contrastive Alignment (KSCA). CAKM formulates cross-modal alignment through hyperspherical Radial Basis Function kernel mapping, optimizing dynamic attribute kernels via prompt learning to capture modality-invariant semantics. Furthermore, KSCA extends conventional point-to-point contrastive learning by modeling limited paired data as continuous kernel distributions. This explicit smoothing of the modality gap alleviates overfitting to sparse pairwise correlations. Extensive experiments demonstrate that APKH outperforms state-of-the-art hashing methods in the challenging cross-modal retrieval tasks from seen to unseen categories under data-constrained scenarios.

Stream guardrails enable token-level safety detection before full responses are generated. However, they often make overly conservative judgements and block those sensitive but safe tokens, which is known as over-refusal. Due to lack of full context, they also fail to detect implicitly harmful content from jailbreaking. To address these challenges, we propose FreoStream, a novel streaming guardrail framework. Specifically, FreoStream fine-tunes a LoRA module to perform Future-Aware Reasoning when the base guardrail detects unsafe tokens. The reasoning process follows a Future-Reason-Judge paradigm: predict the future, reason about the full context and give the final judgement. This design can effectively reduce over-refusal by incorporating the future information. Moreover, we introduce the Safety-Aligned Optimization module that extracts the safety-aligned component from the reasoning gradients to update the base guardrail model, thereby enhancing streaming safety detection. Extensive experiments on various safety benchmarks demonstrate that FreoStream achieves lower over-refusal rates and better jailbreak defense compared to existing streaming guardrails.

Research on robotic manipulation has developed a diverse set of policy paradigms, including vision-language-action (VLA) models, vision-action (VA) policies, and code-based compositional approaches. Concrete policies typically attain high success rates on specific task distributions but lim-ited generalization beyond it. Rather than proposing an other monolithic policy, we propose to leverage the complementary strengths of existing approaches through intelligent policy routing. We introduce RoboRouter, a training-free framework that maintains a pool of heterogeneous policies and learns to select the best-performing policy for each task through accumulated execution experience. Given a new task, RoboRouter constructs a semantic task representation, retrieves historical records of similar tasks, predicts the optimal policy choice without requiring trial-and-error, and incorporates structured feedback to refine subsequent routing decisions. Integrating a new policy into the system requires only lightweight evaluation and incurs no training overhead. Across simulation benchmark and real-world evaluations, RoboRouter consistently outperforms than in-dividual policies, improving average success rate by more than 3

Exemplar-free Class-Incremental Learning (EFCIL) poses a significant challenge in mitigating catastrophic forgetting, due to the absence of exemplars. Recently, analytic learning-based methods propose a recursive alignment procedure to execute EFCIL in a phase-invariant manner and show state-of-the-art performance. However, they heavily rely on a frozen feature extractor trained with the initial dataset to avoid the misalignment between feature and label spaces, ignoring the importance of acquiring generalizable features across incremental tasks for performance improvement. To tackle this, we rethink the obscured sub-optimality of analytic learning-based methods, particularly through empirical reevaluation, and then introduce the Multi-head analytic learning (Muheal) approach. Muheal forms the multi-head model with a delicate feature extractor, thereby introducing a feature optimization procedure and a forgetting compensation module to balance the learning and forgetting. Specifically, within the feature optimization procedure, the feature extractor seeks to learn more generalizable features in a self-supervised manner using the fully-connected classification head. An analytic learning-based classification head follows to align the feature-label space. Additionally, we employ the compensation module to generate and align pseudo-features with a replicated analytic head, thus preventing overfitting and testing. Comprehensive experiments on several benchmark datasets have demonstrated that Muheal significantly outperforms existing state-of-the-art EFCIL methods and is comparable, if not superior, to methods that use replay techniques.

Large Language Models (LLMs) with long chain-of-thought (CoT) capability, termed Reasoning Models, demonstrate superior intricate problem-solving abilities through multi-step long CoT reasoning. To create a dual-capability model with long CoT capability and domain-specific knowledge without substantial computational and data costs, model merging emerges as a highly resource-efficient method. However, significant challenges lie in merging domain-specific LLMs with long CoT ones since nowadays merging methods suffer from reasoning capability degradation, even gibberish output and output collapse. To overcome this, we introduce RCP-Merging: Merging Long Chain-of-Thought Models with Domain-Specific Models by Considering Reasoning Capability as Prior, a novel merging framework designed to integrate domain-specific LLMs with long CoT capability, meanwhile maintaining model performance in the original domain. Treating reasoning model weights as foundational prior, our method utilizes a reasoning capability indicator to preserve core long CoT capability model weights while selectively merging essential domain-specific weights. We conducted extensive experiments on Qwen2.5-7B, Llama3.1-8B, and Qwen2.5-1.5B models in BioMedicine and Finance domains. Our results show that RCP-Merging successfully merges a reasoning model with domain-specific ones, improving domain task performance by 9.5% and 9.2% over state-of-the-art methods, without significantly harming the original long CoT reasoning capability.