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.

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× and achieves 1.52× faster decoding, with almost no performance drop or even higher relative to the full-cache MLLM.

Parallel Speculative Decoding (PSD) accelerates traditional Speculative Decoding (SD) by overlapping draft generation with verification. However, it remains hampered by two fundamental challenges: (1) a theoretical speedup ceiling dictated by the speed ratio between the draft and target models, and (2) high computational waste and pipeline stall due to mid-sequence token rejections of early errors. To address these limitations, we introduce Double (Double Retrieval Speculative Parallelism). By bridging the gap between SD and PSD, our framework resolves the Retrieval Precision-Efficiency Dilemma through a novel synchronous mechanism. Specifically, we enable the draft model to execute iterative retrieval speculations to break the theoretical speedup limits; to alleviate rejections without rollback, the target model performs authoritative retrieval to generate multi-token guidance. Double is entirely training-free and lossless. Extensive experiments demonstrate state-of-the-art speedup of 5.3× on LLaMA3.3-70B and 2.8× on Qwen3-32B, significantly outperforming the advanced method EAGLE-3 that requires extensive model training. Our code is available at https://github.com/Sylvan820/Double1.

Video Large Language Models (Video-LLMs) excel in video understanding but suffer from high inference latency due to 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 is high-fidelity and rapid: it preserves >99.8% of target performance while accelerating Qwen2.5-VL-32B by 2.70 × and LLaVA-OneVision-72B by 2.94 ×. Notably, it boosts the mean accepted length and speedup ratio by 136% and 35% compared to SOTA training-free SD methods for Video-LLMs. Code is provided in the submitted software.

Domain adaptation is widely adopted in text retrieval scenarios where large labeled data is unavailable. To improve model adaptability, existing methods try to expand more source datasets. However, we found from experiments that indiscriminately using a large amount of source data from various text tasks does not guarantee improved adaptability, but may negatively impact model performance. To tackle this issue, we propose Trait, a framework that can effectively improve model adaptability by selecting beneficial data without evaluating all source data. Specifically, we first divide multiple source datasets into data chunks of the same size as the minimum selection unit to form the whole selection space. Then we devise an iterative process that includes Bayesian optimization-based selection and transfer-aware chunk evaluation to incrementally select beneficial chunks. To reduce unnecessary evaluation costs, we also design backtracking and pruning actions to adjust the selection subspace. Extensive experimental results show that Trait not only achieves average state-of-the-art for few-shot on nine target datasets by evaluating only 4% of BERRI source data, but also is very competitive for zero-shot compared with LLM-based rankers.

Video large language models (Vid-LLMs) have shown strong capabilities in understanding video content. However, their reliance on dense video token representations introduces substantial memory and computational overhead in both prefilling and decoding. To mitigate the information loss of recent video token reduction methods and accelerate the decoding stage of Vid-LLMs losslessly, we introduce SpecVLM, a training-free speculative decoding (SD) framework tailored for Vid-LLMs that incorporates staged video token pruning.Building on our novel finding that the draft model’s speculation exhibits low sensitivity to video token pruning, SpecVLM prunes up to 90% of video tokens to enable efficient speculation without sacrificing accuracy. To achieve this, we performs a two-stage pruning process: Stage I selects highly informative tokens guided by attention signals from the verifier (target model), while Stage II prunes remaining redundant ones in a spatially uniform manner.Extensive experiments on four video understanding benchmarks demonstrate the effectiveness and robustness of SpecVLM, which achieves up to 2.68× decoding speedup for LLaVA-OneVision-72B and 2.11× speedup for Qwen2.5-VL-32B. Code is available at https://github.com/zju-jiyicheng/SpecVLM.

Multimodal learning is garnering significant attention for its capacity to represent diverse human perceptions (e.g., linguistic, acoustic, and visual signals), achieving more natural and intuitive interactions with technology.However, the frequent occurrence of incomplete data, either within a single modality (intra-modality) or across different modalities (inter-modality), presents substantial challenges in reliable semantic interpretation and model reasoning.Furthermore, there is currently no robust representation learning mechanism capable of managing both intra-modality and inter-modality real-data deficiencies.To address this challenge, we present T²DR, a two-tier deficiency-resistant framework for incomplete multimodal learning, which comprises two main modules:(1) Intra-Modal Deficiency-Resistant module (IADR): To address fine-grained deficiencies, we introduce Intra-Attn to focus on the available data while avoiding excessive suppression of the missing regions.(2) Inter-Modal Deficiency-Resistant module (IEDR): To handle coarse-grained deficiencies, we propose the shared feature prediction (SFP) to leverage cross-modal shared features for preliminary data imputation. Subsequently, we apply Inter-Attn to allocate appropriate attention to each modality based on the results from the capability-aware scorer (CAS).Extensive experiments are performed on two well-known multimodal benchmarks, CMU-MOSI and CMU-MOSEI, across various missing scenarios for sentiment analysis. Experimental results show that T²DR significantly outperforms the SOTA models. Code is available at https://github.com/LH019/T2DR.

We present a novel inference scheme, self-speculative decoding, for accelerating Large Language Models (LLMs) without the need for an auxiliary model. This approach is characterized by a two-stage process: drafting and verification. The drafting stage generates draft tokens at a slightly lower quality but more quickly, which is achieved by selectively skipping certain intermediate layers during drafting. Subsequently, the verification stage employs the original LLM to validate those draft output tokens in one forward pass. This process ensures the final output remains identical to that produced by the unaltered LLM. Moreover, the proposed method requires no additional neural network training and no extra memory footprint, making it a plug-and-play and cost-effective solution for inference acceleration. Benchmarks with LLaMA-2 and its variants demonstrated a speedup up to 1.99×.