Accurate INT8 Training Through Dynamic Block-Level Fallback

Transformer models have achieved remarkable success across various AI applications but face significant training costs. Low-bit training, such as INT8 training, can leverage computational units with higher throughput, and has already demonstrated its effectiveness on GPT2 models with block-level quantization. However, it struggles with modern Transformer variants incorporating GLU units. This is because those variants demonstrate complex distributions of activation outliers. To address the challenge, we propose Fallback Quantization, implementing mixed-precision GEMM that dynamically falls back 8-bit to 16-bit for activation blocks containing outliers. Experiments show that our approach is robustly competent in both fine-tuning and pretraining settings. Moreover, our method achieves a 1.57x end-to-end training speedup on RTX4090 GPUs.

UAR-NVC: A Unified AutoRegressive Framework for Memory-Efficient Neural Video Compression

Implicit Neural Representations (INRs) have demonstrated significant potential in video compression by representing videos as neural networks. However, as the number of frames increases, the memory consumption for training and inference increases substantially, posing challenges in resource-constrained scenarios. Inspired by the success of traditional video compression frameworks, which process video frame by frame and can efficiently compress long videos, we adopt this modeling strategy for INRs to decrease memory consumption, while aiming to unify the frameworks from the perspective of timeline-based autoregressive modeling. In this work, we present a novel understanding of INR models from an autoregressive (AR) perspective and introduce a Unified AutoRegressive Framework for memory-efficient Neural Video Compression (UAR-NVC). UAR-NVC integrates timeline-based and INR-based neural video compression under a unified autoregressive paradigm. It partitions videos into several clips and processes each clip using a different INR model instance, leveraging the advantages of both compression frameworks while allowing seamless adaptation to either in form. To further reduce temporal redundancy between clips, we design two modules to optimize the initialization, training, and compression of these model parameters. UAR-NVC supports adjustable latencies by varying the clip length. Extensive experimental results demonstrate that UAR-NVC, with its flexible video clip setting, can adapt to resource-constrained environments and significantly improve performance compared to different baseline models.

Direct Discriminative Optimization: Your Likelihood-Based Visual Generative Model is Secretly a GAN Discriminator

While likelihood-based generative models, particularly diffusion and autoregressive models, have achieved remarkable fidelity in visual generation, the maximum likelihood estimation (MLE) objective inherently suffers from a mode-covering tendency that limits the generation quality under limited model capacity. In this work, we propose Direct Discriminative Optimization (DDO) as a unified framework that bridges likelihood-based generative training and the GAN objective to bypass this fundamental constraint. Our key insight is to parameterize a discriminator implicitly using the likelihood ratio between a learnable target model and a fixed reference model, drawing parallels with the philosophy of Direct Preference Optimization (DPO). Unlike GANs, this parameterization eliminates the need for joint training of generator and discriminator networks, allowing for direct, efficient, and effective finetuning of a well-trained model to its full potential beyond the limits of MLE. DDO can be performed iteratively in a self-play manner for progressive model refinement, with each round requiring less than 1% of pretraining epochs. Our experiments demonstrate the effectiveness of DDO by significantly advancing the previous SOTA diffusion model EDM, reducing FID scores from 1.79/1.58 to new records of 1.30/0.97 on CIFAR-10/ImageNet-64 datasets, and by consistently improving both guidance-free and CFG-enhanced FIDs of visual autoregressive models on ImageNet 256$\times$256.

Oscillation-Reduced MXFP4 Training for Vision Transformers

Pre-training Transformers in FP4 precision is becoming a promising approach to gain substantial speedup, but it comes with a considerable loss of accuracy. Microscaling (MX) data format provides a fine-grained per-group quantization method to improve the representation ability of the FP4 format and is supported by the next-generation Blackwell GPU architecture. However, training with MXFP4 data format still results in significant degradation and there is a lack of systematic research on the reason. In this work, we propose a novel training method TetraJet for a more accurate FP4 training. We comprehensively evaluate all of the quantizers involved in the training, and identify the weight oscillation problem in the forward pass as the main source of the degradation in MXFP4 training. Therefore, we introduce two novel methods, EMA Quantizer (Q-EMA) and Adaptive Ramping Optimizer (Q-Ramping), to resolve the oscillation problem. Extensive experiments on Vision Transformers demonstrate that TetraJet consistently outperforms the existing 4-bit training methods, and Q-EMA & Q-Ramping can provide additional enhancement by effectively reducing oscillation. We decreased the accuracy degradation by more than $50\%$ compared to the baseline, and can even achieve competitive performance compared to full precision training. The codes are available at https://github.com/thu-ml/TetraJet-MXFP4Training

SpargeAttn: Accurate Sparse Attention Accelerating Any Model Inference

An efficient attention implementation is essential for large models due to its quadratic time complexity. Fortunately, attention commonly exhibits sparsity, i.e., many values in the attention map are near zero, allowing for the omission of corresponding computations. Many studies have utilized the sparse pattern to accelerate attention. However, most existing works focus on optimizing attention within specific models by exploiting certain sparse patterns of the attention map. A universal sparse attention that guarantees both the speedup and end-to-end performance of diverse models remains elusive. In this paper, we propose SpargeAttn, a universal sparse and quantized attention for any model. Our method uses a two-stage online filter: in the first stage, we rapidly and accurately predict the attention map, enabling the skip of some matrix multiplications in attention. In the second stage, we design an online softmax-aware filter that incurs no extra overhead and further skips some matrix multiplications. Experiments show that our method significantly accelerates diverse models, including language, image, and video generation, without sacrificing end-to-end metrics. The codes are available at https://github.com/thu-ml/SpargeAttn.

RIFLEx: A Free Lunch for Length Extrapolation in Video Diffusion Transformers

Recent advancements in video generation have enabled models to synthesize high-quality, minute-long videos. However, generating even longer videos with temporal coherence remains a major challenge, and existing length extrapolation methods lead to temporal repetition or motion deceleration. In this work, we systematically analyze the role of frequency components in positional embeddings and identify an intrinsic frequency that primarily governs extrapolation behavior. Based on this insight, we propose RIFLEx, a minimal yet effective approach that reduces the intrinsic frequency to suppress repetition while preserving motion consistency, without requiring any additional modifications. RIFLEx offers a true free lunch--achieving high-quality $2\times$ extrapolation on state-of-the-art video diffusion transformers in a completely training-free manner. Moreover, it enhances quality and enables $3\times$ extrapolation by minimal fine-tuning without long videos. Project page and codes: \href{https://riflex-video.github.io/}{https://riflex-video.github.io/.}

Exploratory Diffusion Policy for Unsupervised Reinforcement Learning

Unsupervised reinforcement learning (RL) aims to pre-train agents by exploring states or skills in reward-free environments, facilitating the adaptation to downstream tasks. However, existing methods often overlook the fitting ability of pre-trained policies and struggle to handle the heterogeneous pre-training data, which are crucial for achieving efficient exploration and fast fine-tuning. To address this gap, we propose Exploratory Diffusion Policy (EDP), which leverages the strong expressive ability of diffusion models to fit the explored data, both boosting exploration and obtaining an efficient initialization for downstream tasks. Specifically, we estimate the distribution of collected data in the replay buffer with the diffusion policy and propose a score intrinsic reward, encouraging the agent to explore unseen states. For fine-tuning the pre-trained diffusion policy on downstream tasks, we provide both theoretical analyses and practical algorithms, including an alternating method of Q function optimization and diffusion policy distillation. Extensive experiments demonstrate the effectiveness of EDP in efficient exploration during pre-training and fast adaptation during fine-tuning.

CANeRV: Content Adaptive Neural Representation for Video Compression

Recent advances in video compression introduce implicit neural representation (INR) based methods, which effectively capture global dependencies and characteristics of entire video sequences. Unlike traditional and deep learning based approaches, INR-based methods optimize network parameters from a global perspective, resulting in superior compression potential. However, most current INR methods utilize a fixed and uniform network architecture across all frames, limiting their adaptability to dynamic variations within and between video sequences. This often leads to suboptimal compression outcomes as these methods struggle to capture the distinct nuances and transitions in video content. To overcome these challenges, we propose Content Adaptive Neural Representation for Video Compression (CANeRV), an innovative INR-based video compression network that adaptively conducts structure optimisation based on the specific content of each video sequence. To better capture dynamic information across video sequences, we propose a dynamic sequence-level adjustment (DSA). Furthermore, to enhance the capture of dynamics between frames within a sequence, we implement a dynamic frame-level adjustment (DFA). {Finally, to effectively capture spatial structural information within video frames, thereby enhancing the detail restoration capabilities of CANeRV, we devise a structure level hierarchical structural adaptation (HSA).} Experimental results demonstrate that CANeRV can outperform both H.266/VVC and state-of-the-art INR-based video compression techniques across diverse video datasets.

Effective Black-Box Multi-Faceted Attacks Breach Vision Large Language Model Guardrails

Vision Large Language Models (VLLMs) integrate visual data processing, expanding their real-world applications, but also increasing the risk of generating unsafe responses. In response, leading companies have implemented Multi-Layered safety defenses, including alignment training, safety system prompts, and content moderation. However, their effectiveness against sophisticated adversarial attacks remains largely unexplored. In this paper, we propose MultiFaceted Attack, a novel attack framework designed to systematically bypass Multi-Layered Defenses in VLLMs. It comprises three complementary attack facets: Visual Attack that exploits the multimodal nature of VLLMs to inject toxic system prompts through images; Alignment Breaking Attack that manipulates the model's alignment mechanism to prioritize the generation of contrasting responses; and Adversarial Signature that deceives content moderators by strategically placing misleading information at the end of the response. Extensive evaluations on eight commercial VLLMs in a black-box setting demonstrate that MultiFaceted Attack achieves a 61.56% attack success rate, surpassing state-of-the-art methods by at least 42.18%.

A Regularized Newton Method for Nonconvex Optimization with Global and Local Complexity Guarantees

We consider the problem of finding an $\epsilon$-stationary point of a nonconvex function with a Lipschitz continuous Hessian and propose a quadratic regularized Newton method incorporating a new class of regularizers constructed from the current and previous gradients. The method leverages a recently developed linear conjugate gradient approach with a negative curvature monitor to solve the regularized Newton equation. Notably, our algorithm is adaptive, requiring no prior knowledge of the Lipschitz constant of the Hessian, and achieves a global complexity of $O(\epsilon^{-\frac{3}{2}}) + \tilde O(1)$ in terms of the second-order oracle calls, and $\tilde O(\epsilon^{-\frac{7}{4}})$ for Hessian-vector products, respectively. Moreover, when the iterates converge to a point where the Hessian is positive definite, the method exhibits quadratic local convergence. Preliminary numerical results illustrate the competitiveness of our algorithm.