The performance of Video Question Answering (VideoQA) models is fundamentally constrained by the nature of their supervision, which typically consists of isolated, factual question-answer pairs. This "bag-of-facts" approach fails to capture the underlying narrative and causal structure of events, limiting models to a shallow understanding of video content. To move beyond this paradigm, we introduce a framework to synthesize richer supervisory signals. We propose two complementary strategies: Question-Based Paraphrasing (QBP), which synthesizes the diverse inquiries (what, how, why) from a video's existing set of question-answer pairs into a holistic narrative paragraph that reconstructs the video's event structure; and Question-Based Captioning (QBC), which generates fine-grained visual rationales, grounding the answer to each question in specific, relevant evidence. Leveraging powerful generative models, we use this synthetic data to train VideoQA models under a unified next-token prediction objective. Extensive experiments on STAR and NExT-QA validate our approach, demonstrating significant accuracy gains and establishing new state-of-the-art results, such as improving a 3B model to 72.5\% on STAR (+4.9\%) and a 7B model to 80.8\% on NExT-QA. Beyond accuracy, our analysis reveals that both QBP and QBC substantially enhance cross-dataset generalization, with QBP additionally accelerating model convergence by over 2.5x. These results demonstrate that shifting data synthesis from isolated facts to narrative coherence and grounded rationales yields a more accurate, efficient, and generalizable training paradigm.

This paper presents an evolutionary framework for the training of large language models(LLM). The models are divided into several experts(sub-networks), which have the same structure but different parameter values. Only one expert is trained at each step. After the classical AdamW optimization, some evolutionary operators(crossover, PSO, and mutation) act on the tensor weights between the current expert and the best expert. So current expert would learn the experience of best expert. The direction of best expert would help current expert's loss decrease faster. Finally, only save the weight of the best expert. Experiments show that best expert would achieve nearly the same accuracy as the full model. This would greatly reduce the size of the model for inference. Since only one expert is trained at each step, the training needs much less memory and has much higher throughput. Experiments show that the throughput would accelerate more than ten times! Our source code is available. It's a pure c++/cu framework, which is suitable for easy deployment on PCs and edge computing devices.

In machine learning, research has traditionally focused on model development, with relatively less attention paid to training data. As model architectures have matured and marginal gains from further refinements diminish, data quality has emerged as a critical factor. However, systematic studies on evaluating and ensuring dataset quality in the image domain remain limited. This study investigates methods for systematically assessing image dataset quality and examines how various image quality factors influence model performance. Using the publicly available and relatively clean CIFAKE dataset, we identify common quality issues and quantify their impact on training. Building on these findings, we develop a pipeline that integrates two community-developed tools, CleanVision and Fastdup. We analyze their underlying mechanisms and introduce several enhancements, including automatic threshold selection to detect problematic images without manual tuning. Experimental results demonstrate that not all quality issues exert the same level of impact. While convolutional neural networks show resilience to certain distortions, they are particularly vulnerable to degradations that obscure critical visual features, such as blurring and severe downscaling. To assess the performance of existing tools and the effectiveness of our proposed enhancements, we formulate the detection of low-quality images as a binary classification task and use the F1 score as the evaluation metric. Our automatic thresholding method improves the F1 score from 0.6794 to 0.9468 under single perturbations and from 0.7447 to 0.8557 under dual perturbations. For near-duplicate detection, our deduplication strategy increases the F1 score from 0.4576 to 0.7928. These results underscore the effectiveness of our workflow and provide a foundation for advancing data quality assessment in image-based machine learning.

Visual generation quality has been greatly promoted with the rapid advances in diffusion transformers (DiTs), which is attributed to the scaling of model size and complexity. However, these attributions also hinder the practical deployment of DiTs on edge devices, limiting their development and application. Serve as an efficient model compression technique, model post-training quantization (PTQ) can reduce the memory consumption and speed up the inference, with inevitable performance degradation. To alleviate the degradation, we propose CLQ, a cross-layer guided orthogonal-based quantization method for DiTs. To be specific, CLQ consists of three key designs. First, we observe that the calibration data used by most of the PTQ methods can not honestly represent the distribution of the activations. Therefore, we propose cross-block calibration (CBC) to obtain accurate calibration data, with which the quantization can be better guided. Second, we propose orthogonal-based smoothing (OBS), which quantifies the outlier score of each channel and leverages block Hadamard matrix to smooth the outliers with negligible overhead. Third, we propose cross-layer parameter searching (CLPS) to search. We evaluate CLQ with both image generation and video generation models and successfully compress the model into W4A4 with negligible degradation in visual quality and metrics. CLQ achieves 3.98x memory saving and 3.95x speedup. Our code is available at \hyperlink{https://github.com/Kai-Liu001/CLQ}{https://github.com/Kai-Liu001/CLQ}.

One main challenge in time series anomaly detection for industrial IoT lies in the complex spatio-temporal couplings within multivariate data. However, traditional anomaly detection methods focus on modeling spatial or temporal dependencies independently, resulting in suboptimal representation learning and limited sensitivity to anomalous dispersion in high-dimensional spaces. In this work, we conduct an empirical analysis showing that both normal and anomalous samples tend to scatter in high-dimensional space, especially anomalous samples are markedly more dispersed. We formalize this dispersion phenomenon as scattering, quantified by the mean pairwise distance among sample representations, and leverage it as an inductive signal to enhance spatio-temporal anomaly detection. Technically, we propose ScatterAD to model representation scattering across temporal and topological dimensions. ScatterAD incorporates a topological encoder for capturing graph-structured scattering and a temporal encoder for constraining over-scattering through mean squared error minimization between neighboring time steps. We introduce a contrastive fusion mechanism to ensure the complementarity of the learned temporal and topological representations. Additionally, we theoretically show that maximizing the conditional mutual information between temporal and topological views improves cross-view consistency and enhances more discriminative representations. Extensive experiments on multiple public benchmarks show that ScatterAD achieves state-of-the-art performance on multivariate time series anomaly detection. Code is available at this repository: https://github.com/jk-sounds/ScatterAD.

Spiking Neural Networks (SNNs) have gained significant traction in both computational neuroscience and artificial intelligence for their potential in energy-efficient computing. In contrast, artificial neural networks (ANNs) excel at gradient-based optimization and high accuracy. This contrast has consequently led to a growing subfield of hybrid ANN-SNN research. However, existing hybrid approaches often rely on either a strict separation between ANN and SNN components or employ SNN-only encoders followed by ANN classifiers due to the constraints of non-differentiability of spike encoding functions, causing prior hybrid architectures to lack deep layer-wise cooperation during backpropagation. To address this gap, we propose a novel hybrid ANN-SNN framework that integrates layer-wise encode-decode SNN blocks within conventional ANN pipelines. Central to our method is the use of surrogate gradients for a bit-plane-based spike encoding function, enabling end-to-end differentiable training across ANN and SNN layers. This design achieves competitive accuracy with state-of-the-art pure ANN and SNN models while retaining the potential efficiency and temporal representation benefits of spiking computation. To the best of our knowledge, this is the first implementation of a surrogate gradient for bit plane coding specifically and spike encoder interface in general to be utilized in the context of hybrid ANN-SNN, successfully leading to a new class of hybrid models that pave new directions for future research.

Text-to-SQL enables natural access to databases, yet most benchmarks are English-only, limiting multilingual progress. We introduce MultiSpider 2.0, extending Spider 2.0 to eight languages (English, German, French, Spanish, Portuguese, Japanese, Chinese, Vietnamese). It preserves Spider 2.0's structural difficulty while adding linguistic and dialectal variability, demanding deeper reasoning for complex SQL. On this benchmark, state-of-the-art LLMs (such as DeepSeek-R1 and OpenAI o1) reach only 4\% execution accuracy when relying on intrinsic reasoning, versus 60\% on MultiSpider 1.0. Therefore, we provide a collaboration-driven language agents baseline that iteratively refines queries, improving accuracy to 15\%. These results reveal a substantial multilingual gap and motivate methods that are robust across languages and ready for real-world enterprise deployment. Our benchmark is available at https://github.com/phkhanhtrinh23/Multilingual_Text_to_SQL.

Although Large Reasoning Models (LRMs) have progressed in solving complex problems, their chain-of-thought (CoT) reasoning often contains harmful content that can persist even when the final responses appear safe. We show that this issue still remains in existing methods which overlook the unique significance of safe reasoning, undermining their trustworthiness and posing potential risks in applications if unsafe reasoning is accessible for and exploited by malicious users. We therefore shift our focus to aligning the safety of reasoning itself in this paper and explore process supervision as the solution. However, simply rewarding safe reasoning proves inadequate due to low rollout diversity and limited training signals. To tackle this challenge, we first delve into the characteristics of safe reasoning and uncover several critical insights that 1) safe reasoning is often consolidated by a few critical steps of safety triggers; 2) compliance cues strongly correlate with unsafe continuations; and 3) corrective interventions reliably steer unsafe trajectories towards safer traces. Motivated by these, we propose Intervened Preference Optimization (IPO), an alignment method that enforces safe reasoning by substituting compliance steps with safety triggers and constructing pairs for preference learning with strong signals. Experiments on jailbreak and adversarial safety benchmarks demonstrate that IPO remarkably improves overall safety regarding both reasoning and responses, outperforming SFT-based and RL-based baselines with a relative reduction of over 30% in harmfulness, while preserving excellent performance across diverse reasoning tasks. The results highlight the importance of explicit alignment for reasoning and provide a practical path to safer LRMs.

We introduce LLaDA-MoE, a large language diffusion model with the Mixture-of-Experts (MoE) architecture, trained from scratch on approximately 20T tokens. LLaDA-MoE achieves competitive performance with significantly reduced computational overhead by maintaining a 7B-parameter capacity while activating only 1.4B parameters during inference. Our empirical evaluation reveals that LLaDA-MoE achieves state-of-the-art performance among diffusion language models with larger parameters, surpassing previous diffusion language models LLaDA, LLaDA 1.5, and Dream across multiple benchmarks. The instruct-tuned model LLaDA-MoE-7B-A1B-Instruct demonstrates capabilities comparable to Qwen2.5-3B-Instruct in knowledge understanding, code generation, mathematical reasoning, agent and alignment tasks, despite using fewer active parameters. Our results show that integrating a sparse MoE architecture into the training objective of masked diffusion language models still brings out MoE's strengths under efficient inference with few active parameters, and opens ample room for further exploration of diffusion language models. LLaDA-MoE models are available at Huggingface.

The alignment of large language models (LLMs) with human values is critical for their safe deployment, yet jailbreak attacks can subvert this alignment to elicit harmful outputs from LLMs. In recent years, a proliferation of jailbreak attacks has emerged, accompanied by diverse metrics and judges to assess the harmfulness of the LLM outputs. However, the absence of a systematic benchmark to assess the quality and effectiveness of these metrics and judges undermines the credibility of the reported jailbreak effectiveness and other risks. To address this gap, we introduce HarmMetric Eval, a comprehensive benchmark designed to support both overall and fine-grained evaluation of harmfulness metrics and judges. Our benchmark includes a high-quality dataset of representative harmful prompts paired with diverse harmful and non-harmful model responses, alongside a flexible scoring mechanism compatible with various metrics and judges. With HarmMetric Eval, our extensive experiments uncover a surprising result: two conventional metrics--METEOR and ROUGE-1--outperform LLM-based judges in evaluating the harmfulness of model responses, challenging prevailing beliefs about LLMs' superiority in this domain. Our dataset is publicly available at https://huggingface.co/datasets/qusgo/HarmMetric_Eval, and the code is available at https://anonymous.4open.science/r/HarmMetric-Eval-4CBE.