This paper introduces Completion Pruning Policy Optimization (CPPO) to accelerate the training of reasoning models based on Group Relative Policy Optimization (GRPO). GRPO, while effective, incurs high training costs due to the need for sampling multiple completions for each question. Our experiment and theoretical analysis reveals that the number of completions impacts model accuracy yet increases training time multiplicatively, and not all completions contribute equally to policy training -- their contribution depends on their relative advantage. To address these issues, we propose CPPO, which prunes completions with low absolute advantages, significantly reducing the number needed for gradient calculation and updates. Additionally, we introduce a dynamic completion allocation strategy to maximize GPU utilization by incorporating additional questions, further enhancing training efficiency. Experimental results demonstrate that CPPO achieves up to $8.32\times$ speedup on GSM8K and $3.51\times$ on Math while preserving or even enhancing the accuracy compared to the original GRPO. We release our code at https://github.com/lzhxmu/CPPO.

Large language models (LLMs) excel at processing long sequences, boosting demand for key-value (KV) caching. While recent efforts to evict KV cache have alleviated the inference burden, they often fail to allocate resources rationally across layers with different attention patterns. In this paper, we introduce Cascading and Adaptive KV cache Eviction (CAKE), a novel approach that frames KV cache eviction as a "cake-slicing problem." CAKE assesses layer-specific preferences by considering attention dynamics in both spatial and temporal dimensions, allocates rational cache size for layers accordingly, and manages memory constraints in a cascading manner. This approach enables a global view of cache allocation, adaptively distributing resources across diverse attention mechanisms while maintaining memory budgets. CAKE also employs a new eviction indicator that considers the shifting importance of tokens over time, addressing limitations in existing methods that overlook temporal dynamics. Comprehensive experiments on LongBench and NeedleBench show that CAKE maintains model performance with only 3.2% of the KV cache and consistently outperforms current baselines across various models and memory constraints, particularly in low-memory settings. Additionally, CAKE achieves over 10 speedup in decoding latency compared to full cache when processing contexts of 128K tokens with FlashAttention-2. New models such as GPT-4 (Achiam et al., 2023), Claude 3.5 (Anthropic, 2024), LLaMA 3.1 (Dubey et al., 2024) and Mistral Large 2 (AI, 2024) have extended token processing capacities beyond 128K. Shazeer (2019); Ainslie et al. (2023) partially address this issue by merging key-value heads during the training phase. However, optimizing key-value cache without additional training is crucial for efficient inference of long contexts under memory constraints, particularly in typical deployment scenarios where the model structure is fixed. One way to maintain a manageable KV cache size on the fly is to remove some KV pairs (Xiao et al., 2023; Zhang et al., 2024b; Li et al., 2024b). The idea is to eliminate less important KV pairs based on certain rules. Although recent methods have enhanced pair selection for removal, they typically assign uniform cache sizes across layers, disregarding layer-specific requirements.

We consider the problem of efficiently representing casually captured monocular videos in a spatially- and temporally-coherent manner. While existing approaches predominantly rely on 2D/2.5D techniques treating videos as collections of spatiotemporal pixels, they struggle with complex motions, occlusions, and geometric consistency due to absence of temporal coherence and explicit 3D structure. Drawing inspiration from monocular video as a projection of the dynamic 3D world, we explore representing videos in their intrinsic 3D form through continuous flows of Gaussian primitives in space-time. In this paper, we propose NutWorld, a novel framework that efficiently transforms monocular videos into dynamic 3D Gaussian representations in a single forward pass. At its core, NutWorld introduces a structured spatial-temporal aligned Gaussian (STAG) representation, enabling optimization-free scene modeling with effective depth and flow regularization. Through comprehensive experiments, we demonstrate that NutWorld achieves high-fidelity video reconstruction quality while enabling various downstream applications in real-time. Demos and code will be available at https://github.com/Nut-World/NutWorld.

Scaling laws for inference compute in multi-agent systems remain under-explored compared to single-agent scenarios. This work aims to bridge this gap by investigating the problem of data synthesis through multi-agent sampling, where synthetic responses are generated by sampling from multiple distinct language models. Effective model coordination is crucial for successful multi-agent collaboration. Unlike previous approaches that rely on fixed workflows, we treat model coordination as a multi-step decision-making process, optimizing generation structures dynamically for each input question. We introduce Tree Search-based Orchestrated Agents~(TOA), where the workflow evolves iteratively during the sequential sampling process. To achieve this, we leverage Monte Carlo Tree Search (MCTS), integrating a reward model to provide real-time feedback and accelerate exploration. Our experiments on alignment, machine translation, and mathematical reasoning demonstrate that multi-agent sampling significantly outperforms single-agent sampling as inference compute scales. TOA is the most compute-efficient approach, achieving SOTA performance on WMT and a 71.8\% LC win rate on AlpacaEval. Moreover, fine-tuning with our synthesized alignment data surpasses strong preference learning methods on challenging benchmarks such as Arena-Hard and AlpacaEval.

Diffusion models suffer severe object repetition and local distortion when the inference resolution differs from its pre-trained resolution. We propose AccDiffusion v2, an accurate method for patch-wise higher-resolution diffusion extrapolation without training. Our in-depth analysis in this paper shows that using an identical text prompt for different patches leads to repetitive generation, while the absence of a prompt undermines image details. In response, our AccDiffusion v2 novelly decouples the vanilla image-content-aware prompt into a set of patch-content-aware prompts, each of which serves as a more precise description of a patch. Further analysis reveals that local distortion arises from inaccurate descriptions in prompts about the local structure of higher-resolution images. To address this issue, AccDiffusion v2, for the first time, introduces an auxiliary local structural information through ControlNet during higher-resolution diffusion extrapolation aiming to mitigate the local distortions. Finally, our analysis indicates that global semantic information is conducive to suppressing both repetitive generation and local distortion. Hence, our AccDiffusion v2 further proposes dilated sampling with window interaction for better global semantic information during higher-resolution diffusion extrapolation. We conduct extensive experiments, including both quantitative and qualitative comparisons, to demonstrate the efficacy of our AccDiffusion v2. The quantitative comparison shows that AccDiffusion v2 achieves state-of-the-art performance in image generation extrapolation without training. The qualitative comparison intuitively illustrates that AccDiffusion v2 effectively suppresses the issues of repetitive generation and local distortion in image generation extrapolation. Our code is available at \url{https://github.com/lzhxmu/AccDiffusion_v2}.

In this paper, we propose TextDestroyer, the first training- and annotation-free method for scene text destruction using a pre-trained diffusion model. Existing scene text removal models require complex annotation and retraining, and may leave faint yet recognizable text information, compromising privacy protection and content concealment. TextDestroyer addresses these issues by employing a three-stage hierarchical process to obtain accurate text masks. Our method scrambles text areas in the latent start code using a Gaussian distribution before reconstruction. During the diffusion denoising process, self-attention key and value are referenced from the original latent to restore the compromised background. Latent codes saved at each inversion step are used for replacement during reconstruction, ensuring perfect background restoration. The advantages of TextDestroyer include: (1) it eliminates labor-intensive data annotation and resource-intensive training; (2) it achieves more thorough text destruction, preventing recognizable traces; and (3) it demonstrates better generalization capabilities, performing well on both real-world scenes and generated images.

In an effort to improve the efficiency and scalability of single-image super-resolution (SISR) applications, we introduce AnySR, to rebuild existing arbitrary-scale SR methods into any-scale, any-resource implementation. As a contrast to off-the-shelf methods that solve SR tasks across various scales with the same computing costs, our AnySR innovates in: 1) building arbitrary-scale tasks as any-resource implementation, reducing resource requirements for smaller scales without additional parameters; 2) enhancing any-scale performance in a feature-interweaving fashion, inserting scale pairs into features at regular intervals and ensuring correct feature/scale processing. The efficacy of our AnySR is fully demonstrated by rebuilding most existing arbitrary-scale SISR methods and validating on five popular SISR test datasets. The results show that our AnySR implements SISR tasks in a computing-more-efficient fashion, and performs on par with existing arbitrary-scale SISR methods. For the first time, we realize SISR tasks as not only any-scale in literature, but also as any-resource. Code is available at https://github.com/CrispyFeSo4/AnySR.

Managing long texts is challenging for large language models (LLMs) due to limited context window sizes. This study introduces UIO-LLMs, an unbiased incremental optimization approach for memory-enhanced transformers under long-context settings. We initially conceptualize the process as a streamlined encoder-decoder framework where the weights-shared encoder and decoder respectively encapsulate a context segment into memories and leverage these memories to predict outputs of the subsequent segment. Subsequently, by treating our memory-enhanced transformers as fully-connected recurrent neural networks (RNNs), we refine the training process using the Truncated Backpropagation Through Time (TBPTT) algorithm, which incorporates innovative incremental optimization techniques. These techniques not only diminish time complexity but also address the bias in gradient computation through an unbiased optimization process. UIO-LLMs successfully handle long context, such as extending the context window of Llama2-7b-chat from 4K to 100K tokens with minimal 2% additional parameters, while keeping the inference cost nearly linear as context length increases.

Post-training Sparsity (PTS) is a recently emerged avenue that chases efficient network sparsity with limited data in need. Existing PTS methods, however, undergo significant performance degradation compared with traditional methods that retrain the sparse networks via the whole dataset, especially at high sparsity ratios. In this paper, we attempt to reconcile this disparity by transposing three cardinal factors that profoundly alter the performance of conventional sparsity into the context of PTS. Our endeavors particularly comprise (1) A base-decayed sparsity objective that promotes efficient knowledge transferring from dense network to the sparse counterpart. (2) A reducing-regrowing search algorithm designed to ascertain the optimal sparsity distribution while circumventing overfitting to the small calibration set in PTS. (3) The employment of dynamic sparse training predicated on the preceding aspects, aimed at comprehensively optimizing the sparsity structure while ensuring training stability. Our proposed framework, termed UniPTS, is validated to be much superior to existing PTS methods across extensive benchmarks. As an illustration, it amplifies the performance of POT, a recently proposed recipe, from 3.9% to 68.6% when pruning ResNet-50 at 90% sparsity ratio on ImageNet. We release the code of our paper at https://github.com/xjjxmu/UniPTS.

Multimodal large language models (MLLMs) demand considerable computations for inference due to the extensive parameters and the additional input tokens needed for visual information representation. Herein, we introduce Visual Tokens Withdrawal (VTW), a plug-and-play module to boost MLLMs for rapid inference. Our approach is inspired by two intriguing phenomena we have observed: (1) the attention sink phenomenon that is prevalent in LLMs also persists in MLLMs, suggesting that initial tokens and nearest tokens receive the majority of attention, while middle vision tokens garner minimal attention in deep layers; (2) the presence of information migration, which implies that visual information is transferred to subsequent text tokens within the first few layers of MLLMs. As per our findings, we conclude that vision tokens are not necessary in the deep layers of MLLMs. Thus, we strategically withdraw them at a certain layer, enabling only text tokens to engage in subsequent layers. To pinpoint the ideal layer for vision tokens withdrawal, we initially analyze a limited set of tiny datasets and choose the first layer that meets the Kullback-Leibler divergence criterion. Our VTW approach can cut computational overhead by over 40\% across diverse multimodal tasks while maintaining performance. Our code is released at https://github.com/lzhxmu/VTW.