See the Forest for the Trees: Loosely Speculative Decoding via Visual-Semantic Guidance for Efficient Inference of Video LLMs

Video Large Language Models (Video-LLMs) excel in video understanding but suffer from high inference latency during autoregressive generation. Speculative Decoding (SD) mitigates this by applying a draft-and-verify paradigm, yet existing methods are constrained by rigid exact-match rules, severely limiting the acceleration potential. To bridge this gap, we propose LVSpec, the first training-free loosely SD framework tailored for Video-LLMs. Grounded in the insight that generation is governed by sparse visual-relevant anchors (mandating strictness) amidst abundant visual-irrelevant fillers (permitting loose verification), LVSpec employs a lightweight visual-relevant token identification scheme to accurately pinpoint the former. To further maximize acceptance, we augment this with a position-shift tolerant mechanism that effectively salvages positionally mismatched but semantically equivalent tokens. Experiments demonstrate that LVSpec achieves high fidelity and speed: it preserves >99.8 of target performance while accelerating Qwen2.5-VL-32B by 2.70x and LLaVA-OneVision-72B by 2.94x. Notably, it boosts the mean accepted length and speedup ratio by 136% and 35% compared to SOTA training-free SD methods for Video-LLMs.

HybridKV: Hybrid KV Cache Compression for Efficient Multimodal Large Language Model Inference

Multimodal Large Language Models (MLLMs) have advanced unified reasoning over text, images, and videos, but their inference is hindered by the rapid growth of key-value (KV) caches. Each visual input expands into thousands of tokens, causing caches to scale linearly with context length and remain resident in GPU memory throughout decoding, which leads to prohibitive memory overhead and latency even on high-end GPUs. A common solution is to compress caches under a fixed allocated budget at different granularities: token-level uniformly discards less important tokens, layer-level varies retention across layers, and head-level redistributes budgets across heads. Yet these approaches stop at allocation and overlook the heterogeneous behaviors of attention heads that require distinct compression strategies. We propose HybridKV, a hybrid KV cache compression framework that integrates complementary strategies in three stages: heads are first classified into static or dynamic types using text-centric attention; then a top-down budget allocation scheme hierarchically assigns KV budgets; finally, static heads are compressed by text-prior pruning and dynamic heads by chunk-wise retrieval. Experiments on 11 multimodal benchmarks with Qwen2.5-VL-7B show that HybridKV reduces KV cache memory by up to $7.9\times$ and achieves $1.52\times$ faster decoding, with almost no performance drop or even higher relative to the full-cache MLLM.

Efficient Inference for Large Vision-Language Models: Bottlenecks, Techniques, and Prospects

Large Vision-Language Models (LVLMs) enable sophisticated reasoning over images and videos, yet their inference is hindered by a systemic efficiency barrier known as visual token dominance. This overhead is driven by a multi-regime interplay between high-resolution feature extraction, quadratic attention scaling, and memory bandwidth constraints. We present a systematic taxonomy of efficiency techniques structured around the inference lifecycle, consisting of encoding, prefilling, and decoding. Unlike prior reviews focused on isolated optimizations, we analyze the end-to-end pipeline to reveal how upstream decisions dictate downstream bottlenecks, covering compute-bound visual encoding, the intensive prefilling of massive contexts, and the ''visual memory wall'' in bandwidth-bound decoding. By decoupling the efficiency landscape into the axes of shaping information density, managing long-context attention, and overcoming memory limits, this work provides a structured analysis of how isolated optimizations compose to navigate the trade-off between visual fidelity and system efficiency. The survey concludes by outlining four future frontiers supported by pilot empirical insights, including hybrid compression based on functional unit sensitivity, modality-aware decoding with relaxed verification, progressive state management for streaming continuity, and stage-disaggregated serving through hardware-algorithm co-design. The submitted software contains a snapshot of our literature repository, which is designed to be maintained as a living resource for the community.

Bridging Pixels and Words: Mask-Aware Local Semantic Fusion for Multimodal Media Verification

As multimodal misinformation becomes more sophisticated, its detection and grounding are crucial. However, current multimodal verification methods, relying on passive holistic fusion, struggle with sophisticated misinformation. Due to 'feature dilution,' global alignments tend to average out subtle local semantic inconsistencies, effectively masking the very conflicts they are designed to find. We introduce MaLSF (Mask-aware Local Semantic Fusion), a novel framework that shifts the paradigm to active, bidirectional verification, mimicking human cognitive cross-referencing. MaLSF utilizes mask-label pairs as semantic anchors to bridge pixels and words. Its core mechanism features two innovations: 1) a Bidirectional Cross-modal Verification (BCV) module that acts as an interrogator, using parallel query streams (Text-as-Query and Image-as-Query) to explicitly pinpoint conflicts; and 2) a Hierarchical Semantic Aggregation (HSA) module that intelligently aggregates these multi-granularity conflict signals for task-specific reasoning. In addition, to extract fine-grained mask-label pairs, we introduce a set of diverse mask-label pair extraction parsers. MaLSF achieves state-of-the-art performance on both the DGM4 and multimodal fake news detection tasks. Extensive ablation studies and visualization results further verify its effectiveness and interpretability.

ParallelVLM: Lossless Video-LLM Acceleration with Visual Alignment Aware Parallel Speculative Decoding

Although current Video-LLMs achieve impressive performance in video understanding tasks, their autoregressive decoding efficiency remains constrained by the massive number of video tokens. Visual token pruning can partially ease this bottleneck, yet existing approaches still suffer from information loss and yield only modest acceleration in decoding. In this paper, we propose ParallelVLM, a training-free draft-then-verify speculative decoding framework that overcomes both mutual waiting and limited speedup-ratio problems between draft and target models in long-video settings. ParallelVLM features two parallelized stages that maximize hardware utilization and incorporate an Unbiased Verifier-Guided Pruning strategy to better align the draft and target models by eliminating the positional bias in attention-guided pruning. Extensive experiments demonstrate that ParallelVLM effectively expands the draft window by $1.6\sim1.8\times$ with high accepted lengths, and accelerates various video understanding benchmarks by 3.36$\times$ on LLaVA-Onevision-72B and 2.42$\times$ on Qwen2.5-VL-32B compared with vanilla autoregressive decoding.

MeMix: Writing Less, Remembering More for Streaming 3D Reconstruction

Reconstruction is a fundamental task in 3D vision and a fundamental capability for spatial intelligence. Particularly, streaming 3D reconstruction is central to real-time spatial perception, yet existing recurrent online models often suffer from progressive degradation on long sequences due to state drift and forgetting, motivating inference-time remedies. We present MeMix, a training-free, plug-and-play module that improves streaming reconstruction by recasting the recurrent state into a Memory Mixture. MeMix partitions the state into multiple independent memory patches and updates only the least-aligned memory patches while exactly preserving others. This selective update mitigates catastrophic forgetting while retaining $O(1)$ inference memory, and requires no fine-tuning or additional learnable parameters, making it directly applicable to existing recurrent reconstruction models. Across standard benchmarks (ScanNet, 7-Scenes, KITTI, etc.), under identical backbones and inference settings, MeMix reduces reconstruction completeness error by 15.3% on average (up to 40.0%) across 300--500 frame streams on 7-Scenes. The code is available at https://dongjiacheng06.github.io/MeMix/

VERA: Identifying and Leveraging Visual Evidence Retrieval Heads in Long-Context Understanding

While Vision-Language Models (VLMs) have shown promise in textual understanding, they face significant challenges when handling long context and complex reasoning tasks. In this paper, we dissect the internal mechanisms governing long-context processing in VLMs to understand their performance bottlenecks. Through the lens of attention analysis, we identify specific Visual Evidence Retrieval (VER) Heads - a sparse, dynamic set of attention heads critical for locating visual cues during reasoning, distinct from static OCR heads. We demonstrate that these heads are causal to model performance; masking them leads to significant degradation. Leveraging this discovery, we propose VERA (Visual Evidence Retrieval Augmentation), a training-free framework that detects model uncertainty (i.e., entropy) to trigger the explicit verbalization of visual evidence attended by VER heads. Comprehensive experiments demonstrate that VERA significantly improves long-context understanding of open-source VLMs: it yields an average relative improvement of 21.3% on Qwen3-VL-8B-Instruct and 20.1% on GLM-4.1V-Thinking across five benchmarks.

Variance-Adaptive Muon: Accelerating LLM Pretraining with NSR-Modulated and Variance-Scaled Momentum

Large Language Models (LLMs) achieve competitive performance across diverse natural language processing (NLP) tasks, yet pretraining is computationally demanding, making optimizer efficiency an important practical consideration. Muon accelerates LLM pretraining via orthogonal momentum updates that serve as a matrix analogue of the element-wise sign operator. Motivated by the recent perspective that Adam is a variance-adaptive sign update algorithm, we propose two variants of Muon, Muon-NSR and Muon-VS, which apply variance-adaptive normalization to momentum before orthogonalization. Muon-NSR applies noise-to-signal ratio (NSR) modulation, while Muon-VS performs variance-based scaling without introducing additional hyperparameters. Experiments on GPT-2 and LLaMA pretraining demonstrate that our proposed methods accelerate convergence and consistently achieve lower validation loss than both competitive, well-tuned AdamW and Muon baselines. For example, on the LLaMA-1.2B model, Muon-NSR and Muon-VS reduce the iterations required to reach the target validation loss by $1.36\times$ relative to the well-tuned Muon following the recent benchmark.

Convergence Rate Analysis of the AdamW-Style Shampoo: Unifying One-sided and Two-Sided Preconditioning

This paper studies the AdamW-style Shampoo optimizer, an effective implementation of classical Shampoo that notably won the external tuning track of the AlgoPerf neural network training algorithm competition. Our analysis unifies one-sided and two-sided preconditioning and establishes the convergence rate $\frac{1}{K}\sum_{k=1}^K E\left[\|\nabla f(X_k)\|_*\right]\leq O(\frac{\sqrt{m+n}C}{K^{1/4}})$ measured by nuclear norm, where $K$ represents the iteration number, $(m,n)$ denotes the size of matrix parameters, and $C$ matches the constant in the optimal convergence rate of SGD. Theoretically, we have $\|\nabla f(X)\|_F\leq \|\nabla f(X)\|_*\leq \sqrt{m+n}\|\nabla f(X)\|_F$, supporting that our convergence rate can be considered to be analogous to the optimal $\frac{1}{K}\sum_{k=1}^KE\left[\|\nabla f(X_k)\|_F\right]\leq O(\frac{C}{K^{1/4}})$ convergence rate of SGD in the ideal case of $\|\nabla f(X)\|_*= Θ(\sqrt{m+n})\|\nabla f(X)\|_F$.

SafeLoad: Efficient Admission Control Framework for Identifying Memory-Overloading Queries in Cloud Data Warehouses

Memory overload is a common form of resource exhaustion in cloud data warehouses. When database queries fail due to memory overload, it not only wastes critical resources such as CPU time but also disrupts the execution of core business processes, as memory-overloading (MO) queries are typically part of complex workflows. If such queries are identified in advance and scheduled to memory-rich serverless clusters, it can prevent resource wastage and query execution failure. Therefore, cloud data warehouses desire an admission control framework with high prediction precision, interpretability, efficiency, and adaptability to effectively identify MO queries. However, existing admission control frameworks primarily focus on scenarios like SLA satisfaction and resource isolation, with limited precision in identifying MO queries. Moreover, there is a lack of publicly available MO-labeled datasets with workloads for training and benchmarking. To tackle these challenges, we propose SafeLoad, the first query admission control framework specifically designed to identify MO queries. Alongside, we release SafeBench, an open-source, industrial-scale benchmark for this task, which includes 150 million real queries. SafeLoad first filters out memory-safe queries using the interpretable discriminative rule. It then applies a hybrid architecture that integrates both a global model and cluster-level models, supplemented by a misprediction correction module to identify MO queries. Additionally, a self-tuning quota management mechanism dynamically adjusts prediction quotas per cluster to improve precision. Experimental results show that SafeLoad achieves state-of-the-art prediction performance with low online and offline time overhead. Specifically, SafeLoad improves precision by up to 66% over the best baseline and reduces wasted CPU time by up to 8.09x compared to scenarios without SafeLoad.