Rethinking Practical and Efficient Quantization Calibration for Vision-Language Models

Post-training quantization (PTQ) is a primary approach for deploying large language models without fine-tuning, and the quantized performance is often strongly affected by the calibration in PTQ. By contrast, in vision-language models (VLMs), substantial differences between visual and text tokens in their activation distributions and sensitivities to quantization error pose significant challenges for effective calibration during PTQ. In this work, we rethink what PTQ calibration should align with in VLMs and propose the Token-level Importance-aware Layer-wise Quantization framework (TLQ). Guided by gradient information, we design a token-level importance integration mechanism for quantization error, and use it to construct a token-level calibration set, enabling a more fine-grained calibration strategy. Furthermore, TLQ introduces a multi-GPU, quantization-exposed layer-wise calibration scheme. This scheme keeps the layer-wise calibration procedure consistent with the true quantized inference path and distributes the complex layer-wise calibration workload across multiple RTX3090 GPUs, thereby reducing reliance on the large memory of A100 GPUs. TLQ is evaluated across two models, three model scales, and two quantization settings, consistently achieving performance improvements across all settings, indicating its strong quantization stability. The code will be released publicly.

Open-Text Aerial Detection: A Unified Framework For Aerial Visual Grounding And Detection

Open-Vocabulary Aerial Detection (OVAD) and Remote Sensing Visual Grounding (RSVG) have emerged as two key paradigms for aerial scene understanding. However, each paradigm suffers from inherent limitations when operating in isolation: OVAD is restricted to coarse category-level semantics, while RSVG is structurally limited to single-target localization. These limitations prevent existing methods from simultaneously supporting rich semantic understanding and multi-target detection. To address this, we propose OTA-Det, the first unified framework that bridges both paradigms into a cohesive architecture. Specifically, we introduce a task reformulation strategy that unifies task objectives and supervision mechanisms, enabling joint training across datasets from both paradigms with dense supervision signals. Furthermore, we propose a dense semantic alignment strategy that establishes explicit correspondence at multiple granularities, from holistic expressions to individual attributes, enabling fine-grained semantic understanding. To ensure real-time efficiency, OTA-Det builds upon the RT-DETR architecture, extending it from closed-set detection to open-text detection by introducing several high efficient modules, achieving state-of-the-art performance on six benchmarks spanning both OVAD and RSVG tasks while maintaining real-time inference at 34 FPS.

PIO-FVLM: Rethinking Training-Free Visual Token Reduction for VLM Acceleration from an Inference-Objective Perspective

Recently, reducing redundant visual tokens in vision-language models (VLMs) to accelerate VLM inference has emerged as a hot topic. However, most existing methods rely on heuristics constructed based on inter-visual-token similarity or cross-modal visual-text similarity, which gives rise to certain limitations in compression performance and practical deployment. In contrast, we propose PIO-FVLM from the perspective of inference objectives, which transforms visual token compression into preserving output result invariance and selects tokens primarily by their importance to this goal. Specially, vision tokens are reordered with the guidance of token-level gradient saliency generated by our designed layer-local proxy loss, a coarse constraint from the current layer to the final result. Then the most valuable vision tokens are selected following the non-maximum suppression (NMS) principle. The proposed PIO-FVLM is training-free and compatible with FlashAttention, friendly to practical application and deployment. It can be deployed independently as an encoder-free method, or combined with encoder compression approaches like VisionZip for use as an encoder-involved method. On LLaVA-Next-7B, PIO-FVLM retains just 11.1% of visual tokens but maintains 97.2% of the original performance, with a 2.67$\times$ prefill speedup, 2.11$\times$ inference speedup, 6.22$\times$ lower FLOPs, and 6.05$\times$ reduced KV Cache overhead. Our code is available at https://github.com/ocy1/PIO-FVLM.

Delving into Muon and Beyond: Deep Analysis and Extensions

The Muon optimizer has recently attracted considerable attention for its strong empirical performance and use of orthogonalized updates on matrix-shaped parameters, yet its underlying mechanisms and relationship to adaptive optimizers such as Adam remain insufficiently understood. In this work, we aim to address these questions through a unified spectral perspective. Specifically, we view Muon as the p = 0 endpoint of a family of spectral transformations of the form U \boldsymbolΣ^{p} V' , and consider additional variants with p = 1/2 , p = 1/4 , and p = 1 . These transformations are applied to both first-moment updates, as in momentum SGD, and to root-mean-square (RMS) normalized gradient updates as in Adam. To enable efficient computation, we develop a coupled Newton iteration that avoids explicit singular value decomposition. Across controlled experiments, we find that RMS-normalized updates yield more stable optimization than first-moment updates. Moreover, while spectral compression provides strong stabilization benefits under first-moment updates, the Muon update (p = 0) does not consistently outperform Adam. These results suggest that Muon is best understood as an effective form of spectral normalization, but not a universally superior optimization method. Our source code will be released at https://github.com/Ocram7/BeyondMuon.

SimpleGPT: Improving GPT via A Simple Normalization Strategy

In this work, we revisit Transformer optimization through the lens of second-order geometry and establish a direct connection between architectural design, activation scale, the Hessian matrix, and the maximum tolerable learning rate. We introduce a simple normalization strategy, termed SimpleNorm, which stabilizes intermediate activation scales by construction. Then, by analyzing the Hessian of the loss with respect to network activations, we theoretically show that SimpleNorm significantly reduces the spectral norm of the Hessian, thereby permitting larger stable learning rates. We validate our theoretical findings through extensive experiments on large GPT models at parameter scales 1B, 1.4B, 7B and 8B. Empirically, SimpleGPT, our SimpleNorm-based network, tolerates learning rates 3$\times$-10$\times$ larger than standard convention, consistently demonstrates strong optimization stability, and achieves substantially better performance than well-established baselines. Specifically, when training 7B-scale models for 60K steps, SimpleGPT achieves a training loss that is 0.08 lower than that of LLaMA2 with QKNorm, reducing the loss from 2.290 to 2.208. Our source code will be released at https://github.com/Ocram7/SimpleGPT.

LLM-I2I: Boost Your Small Item2Item Recommendation Model with Large Language Model

Item-to-Item (I2I) recommendation models are widely used in real-world systems due to their scalability, real-time capabilities, and high recommendation quality. Research to enhance I2I performance focuses on two directions: 1) model-centric approaches, which adopt deeper architectures but risk increased computational costs and deployment complexity, and 2) data-centric methods, which refine training data without altering models, offering cost-effectiveness but struggling with data sparsity and noise. To address these challenges, we propose LLM-I2I, a data-centric framework leveraging Large Language Models (LLMs) to mitigate data quality issues. LLM-I2I includes (1) an LLM-based generator that synthesizes user-item interactions for long-tail items, alleviating data sparsity, and (2) an LLM-based discriminator that filters noisy interactions from real and synthetic data. The refined data is then fused to train I2I models. Evaluated on industry (AEDS) and academic (ARD) datasets, LLM-I2I consistently improves recommendation accuracy, particularly for long-tail items. Deployed on a large-scale cross-border e-commerce platform, it boosts recall number (RN) by 6.02% and gross merchandise value (GMV) by 1.22% over existing I2I models. This work highlights the potential of LLMs in enhancing data-centric recommendation systems without modifying model architectures.

UVLM: Benchmarking Video Language Model for Underwater World Understanding

Recently, the remarkable success of large language models (LLMs) has achieved a profound impact on the field of artificial intelligence. Numerous advanced works based on LLMs have been proposed and applied in various scenarios. Among them, video language models (VidLMs) are particularly widely used. However, existing works primarily focus on terrestrial scenarios, overlooking the highly demanding application needs of underwater observation. To overcome this gap, we introduce UVLM, an under water observation benchmark which is build through a collaborative approach combining human expertise and AI models. To ensure data quality, we have conducted in-depth considerations from multiple perspectives. First, to address the unique challenges of underwater environments, we selected videos that represent typical underwater challenges including light variations, water turbidity, and diverse viewing angles to construct the dataset. Second, to ensure data diversity, the dataset covers a wide range of frame rates, resolutions, 419 classes of marine animals, and various static plants and terrains. Next, for task diversity, we adopted a structured design where observation targets are categorized into two major classes: biological and environmental. Each category includes content observation and change/action observation, totaling 20 distinct task types. Finally, we designed several challenging evaluation metrics to enable quantitative comparison and analysis of different methods. Experiments on two representative VidLMs demonstrate that fine-tuning VidLMs on UVLM significantly improves underwater world understanding while also showing potential for slight improvements on existing in-air VidLM benchmarks, such as VideoMME and Perception text. The dataset and prompt engineering will be released publicly.

Language Embedding Meets Dynamic Graph: A New Exploration for Neural Architecture Representation Learning

Neural Architecture Representation Learning aims to transform network models into feature representations for predicting network attributes, playing a crucial role in deploying and designing networks for real-world applications. Recently, inspired by the success of transformers, transformer-based models integrated with Graph Neural Networks (GNNs) have achieved significant progress in representation learning. However, current methods still have some limitations. First, existing methods overlook hardware attribute information, which conflicts with the current trend of diversified deep learning hardware and limits the practical applicability of models. Second, current encoding approaches rely on static adjacency matrices to represent topological structures, failing to capture the structural differences between computational nodes, which ultimately compromises encoding effectiveness. In this paper, we introduce LeDG-Former, an innovative framework that addresses these limitations through the synergistic integration of language-based semantic embedding and dynamic graph representation learning. Specifically, inspired by large language models (LLMs), we propose a language embedding framework where both neural architectures and hardware platform specifications are projected into a unified semantic space through tokenization and LLM processing, enabling zero-shot prediction across different hardware platforms for the first time. Then, we propose a dynamic graph-based transformer for modeling neural architectures, resulting in improved neural architecture modeling performance. On the NNLQP benchmark, LeDG-Former surpasses previous methods, establishing a new SOTA while demonstrating the first successful cross-hardware latency prediction capability. Furthermore, our framework achieves superior performance on the cell-structured NAS-Bench-101 and NAS-Bench-201 datasets.

MiniMax-Remover: Taming Bad Noise Helps Video Object Removal

Recent advances in video diffusion models have driven rapid progress in video editing techniques. However, video object removal, a critical subtask of video editing, remains challenging due to issues such as hallucinated objects and visual artifacts. Furthermore, existing methods often rely on computationally expensive sampling procedures and classifier-free guidance (CFG), resulting in slow inference. To address these limitations, we propose MiniMax-Remover, a novel two-stage video object removal approach. Motivated by the observation that text condition is not best suited for this task, we simplify the pretrained video generation model by removing textual input and cross-attention layers, resulting in a more lightweight and efficient model architecture in the first stage. In the second stage, we distilled our remover on successful videos produced by the stage-1 model and curated by human annotators, using a minimax optimization strategy to further improve editing quality and inference speed. Specifically, the inner maximization identifies adversarial input noise ("bad noise") that makes failure removals, while the outer minimization step trains the model to generate high-quality removal results even under such challenging conditions. As a result, our method achieves a state-of-the-art video object removal results with as few as 6 sampling steps and doesn't rely on CFG, significantly improving inference efficiency. Extensive experiments demonstrate the effectiveness and superiority of MiniMax-Remover compared to existing methods. Codes and Videos are available at: https://minimax-remover.github.io.

Taming Transformer Without Using Learning Rate Warmup

Scaling Transformer to a large scale without using some technical tricks such as learning rate warump and using an obviously lower learning rate is an extremely challenging task, and is increasingly gaining more attention. In this paper, we provide a theoretical analysis for the process of training Transformer and reveal the rationale behind the model crash phenomenon in the training process, termed \textit{spectral energy concentration} of ${\bW_q}^{\top} \bW_k$, which is the reason for a malignant entropy collapse, where ${\bW_q}$ and $\bW_k$ are the projection matrices for the query and the key in Transformer, respectively. To remedy this problem, motivated by \textit{Weyl's Inequality}, we present a novel optimization strategy, \ie, making the weight updating in successive steps smooth -- if the ratio $\frac{\sigma_{1}(\nabla \bW_t)}{\sigma_{1}(\bW_{t-1})}$ is larger than a threshold, we will automatically bound the learning rate to a weighted multiple of $\frac{\sigma_{1}(\bW_{t-1})}{\sigma_{1}(\nabla \bW_t)}$, where $\nabla \bW_t$ is the updating quantity in step $t$. Such an optimization strategy can prevent spectral energy concentration to only a few directions, and thus can avoid malignant entropy collapse which will trigger the model crash. We conduct extensive experiments using ViT, Swin-Transformer and GPT, showing that our optimization strategy can effectively and stably train these Transformers without using learning rate warmup.