Recent advancements in audio diffusion models have significantly improved text-to-audio editing via inversion techniques. However, these models typically rely on dense, fixed-step sampling trajectories to maintain structural integrity during inversion and generation, leading to prohibitive computational costs. We propose AdaTE, a model-agnostic Adaptive Trajectory Extrapolation framework that accelerates the inversion-based editing process by dynamically evaluating only the most critical generative phases. Specifically, we introduce a hierarchical probing mechanism that monitors curvature acceleration and information gain to detect pivotal transitions within the latent flow. This allows the model to selectively skip redundant segments via linear extrapolation while preserving dense neural evaluations for complex semantic changes. Extensive experiments across AudioLDM2, Auffusion, and Tango2 demonstrate that AdaTE achieves up to a 3.9× speedup with negligible loss in fidelity. AdaTE significantly shifts the Pareto frontier, providing an efficient solution for high-fidelity audio synthesis and editing.

Prefilling computational costs pose a significant bottleneck for Large Language Models (LLMs) and Large Multimodal Models (LMMs) in long-context settings. While token pruning reduces sequence length, prior methods rely on heuristics that break compatibility with hardware-efficient kernels like FlashAttention. In this work, we observe that tokens evolve toward semantic fixing points, making further processing redundant. To this end, we introduce Delta Attention Selective Halting (DASH), a training-free policy that monitors the layer-wise update dynamics of the self-attention mechanism to selectively halt stabilized tokens. Extensive evaluation confirms that DASH generalizes across language and vision benchmarks, delivering significant prefill speedups while preserving model accuracy and hardware efficiency. Code will be released at https://github.com/verach3n/DASH.git .

Diffusion large language models (dLLMs) present a promising alternative to dominant autoregressive models (ARMs) by the ability of parallel decoding at the expense of substantial computation and memory costs. Specifically, the cache mechanism for bidirectional attention in dLLMs demands large memory footprint, restricting their ability to handle long contexts under resource-limited settings. Existing cache eviction strategies are primarily designed for ARMs and fail to account for the role of mask tokens and specific characteristics in dLLMs, resulting in suboptimal performance.To address these challenges, we introduce MaskKV, a training-free cache eviction framework tailored to dLLMs, focusing on the effect of mask tokens in dLLMs. MaskKV is built on two key innovations: (1) a mask-query guided scoring mechanism that leverages attention weights to identify and evict less critical prompt tokens for each head; (2) an adaptive cache budgeting strategy that improves efficiency by reducing allocation in intermediate layers and concentrating resources on prompt-preferring heads. On LLaDA with MaskKV, compressing the KV cache to only 256 pairs (less than 5% of tokens) retains 94% of the full-cache performance on LongBench and achieves up to 31 × acceleration at 32k prompt length. Our code will be released on Github.

Recent efforts to accelerate inference in Multimodal Large Language Models (MLLMs) have largely focused on visual token compression. The effectiveness of these methods is commonly evaluated by measuring the accuracy drop on existing MLLM benchmarks before and after compression. However, these benchmarks are originally designed to assess general perception and reasoning abilities, rather than the specific challenges posed by visual token compression, leading to a fundamental task mismatch. In this work, we uncover a counterintuitive yet consistent phenomenon: simple image downsampling outperforms many advanced visual token compression methods across multiple widely used benchmarks. Through a comprehensive empirical study spanning eight popular benchmarks and multiple state-of-the-art compression techniques, we show that (i) current benchmarks contain substantial noise (task-irrelevant samples) for evaluating visual token compression, and (ii) downsampling can act as an effective data filter that distinguishes between simple and difficult samples with respect to compression sensitivity. Motivated by these findings, we propose VTC-Bench, an evaluation framework that explicitly leverages downsampling as a discriminator to denoise existing benchmarks, enabling a fairer and more meaningful additional assessment of visual token compression methods.

Vision agent memory has shown remarkable effectiveness in long-video understanding; however, storing such memory for videos incurs substantial overhead, leading to high costs in both storage and computation. To address this issue, we propose StreamMeCo, an efficient Stream Agent Memory Compression framework. Specifically, based on the connectivity of the memory graph, StreamMeCo introduces edge-free minmax sampling for isolated nodes and edge-aware weight pruning for connected nodes, evicting redundant memory nodes while maintaining accuracy. In addition, we introduce a time-decay memory retrieval mechanism to mitigate the performance degradation caused by memory compression. Extensive experiments on three challenging benchmark datasets (M3-Bench-robot, M3-Bench-web, and Video-MME-Long) demonstrate that under 70% memory graph compression, StreamMeCo achieves a 1.87× speedup in memory retrieval while delivering an average accuracy improvement of 1.0%. Our code is available at https://github.com/Celina-love-sweet/StreamMeCo.

Large Language Model-based Multi-Agent Systems (MAS) have demonstrated remarkable capabilities in complex tasks. However, manually designing optimal communication topologies is labor-intensive, while automated expansion methods often result in bloated structures with redundant agents, leading to excessive token consumption. To address this problem, we introduce AgentSlimming, a plug-and-play compression framework for graph-structured multi-agent workflows. Motivated by the AgentPruner and AgentQuant in neural networks, AgentSlimming compresses workflows by firstly estimate the importance score of each agent with a hybrid mechanism, and then removing redundant agents or replacing them with low-cost ones, where each operation is then validated with a baseline-anchored acceptance rule to prevent performance collapse. Experiments show that AgentSlimming reduces average token cost by up to 78.9% with negligible performance degradation, and even sometimes improves accuracy, achieving a strong Pareto-optimal trade-off between cost and quality.

Autoregressive (AR) language modeling remains the dominant paradigm due to its dense supervision signal and highly optimized serving infrastructure, but its strictly causal, token-by-token decoding limits parallelism and non-causal modeling. While masked diffusion offers a promising path toward parallel generation, it faces two critical bottlenecks: training inefficiency stemming from sparse masked objectives, and high latency caused by iterative whole-sequence denoising. We present a systematic study of blockwise discrete diffusion, a pragmatic middle ground that preserves AR-compatible serving while enabling parallel intra-block generation. Our study proceeds in four steps: (i) a controlled, compute- and scale-matched comparison revealing that AR is a more effective backbone for blockwise hybrids than masked diffusion objectives; (ii) a scalable conversion recipe, SDAR, validating that AR models spanning 1.7B to 30B parameters can be adapted into block diffusion models with minimal compute while preserving backbone capabilities; and (iii) a systematic characterization of decoding dynamics, which reveals a virtuous cycle where larger models enable more aggressive parallel decoding, achieving theoretical speedups over 5× and wall-clock speedups of 2.3× on H200 GPUs in latency-critical regimes; and (iv) an investigation of local non-causal modeling capabilities, showing that SDAR’s local bidirectional attention overcomes causal bottlenecks in scientific domains (e.g., chemistry) and enables robust test-time scaling. We release the full model suite, the training framework, and our inference engines for further innovation in non-autoregressive generative paradigms.

Block-wise discrete diffusion offers an attractive balance between parallel generation and causal dependency modeling, making it a promising backbone for vision-language modeling. However, its practical adoption has been limited by high training cost, slow convergence, and instability, which have so far kept it behind strong autoregressive (AR) baselines. We present SDAR-VL, the first systematic application of block-wise discrete diffusion to large-scale vision-language understanding (VLU), together with an integrated framework for efficient and stable training. This framework unifies three components: 1) Asynchronous Block-wise Noise Scheduling to diversify supervision within each batch; 2) Effective Mask Ratio Scaling for unbiased loss normalization under stochastic masking; and 3) a Progressive Beta Noise Curriculum that increases effective mask coverage while preserving corruption diversity. Experiments on 21 single-image, multi-image, and video benchmarks show that SDAR-VL consistently improves training efficiency, convergence stability, and task performance over conventional block diffusion. On this evaluation suite, SDAR-VL sets a new state of the art among diffusion-based vision-language models and, under matched settings, matches or surpasses strong AR baselines such as LLaVA-OneVision as well as the global diffusion baseline LLaDA-V, establishing block-wise diffusion as a practical backbone for VLU.

Dataset Pruning (DP) aims to construct a coreset that achieves performance comparable to the original, full dataset. However, few studies have explored DP in the context of Speech Classification (SC) tasks. Unlike image or text classification, SC is particularly challenging due to the difficulty in capturing the acoustic, semantic, and contextual representations. In this study, we propose a novel dataset pruning method for speech datasets, termed Meltrim, which uses a two-step coarse-to-fine framework designed to address these challenges. Specifically, in Step 1, Meltrim coarsely filters utterance-level redundant samples using DBSCAN clustering on Mel-Frequency Cepstral Coefficients (MFCC) features, which are first flattened and then reduced in dimensionality using UMAP. In Step 2, we perform frame-level redundancy pruning for each utterance via utility pruning, which aims to eliminate irrelevant frames within each utterance. To the best of our knowledge, this is the first dataset pruning approach designed for Speech Classification tasks, demonstrating outstanding performance compared to classical general DP methods. Notably, for the Speech Emotion Recognition, our method achieves up to a 49.5% improvement in WA (Weighted Accuracy) on the MEAD dataset. For the Speaker Identification tasks, it results in a 41.9% reduction in EER (Equal Error Rate) on the VoxCeleb1 dataset.

End-to-end Large Speech Language Models (**LSLMs**) demonstrate strong potential in response latency and speech comprehension capabilities, showcasing general intelligence across speech understanding tasks. However, the ability to follow speech instructions has not been fully realized due to the lack of datasets and heavily biased training tasks. Leveraging the rich ASR datasets, previous approaches have used Large Language Models (**LLMs**) to continue the linguistic information of speech to construct speech instruction datasets. Yet, due to the gap between LLM-generated results and real human responses, the continuation methods further amplify these shortcomings. Given the high costs of collecting and annotating speech instruction datasets by humans, using speech synthesis to construct large-scale speech instruction datasets has become a balanced and robust alternative. Although modern Text-To-Speech (**TTS**) models have achieved near-human-level synthesis quality, it is challenging to appropriately convert out-of-distribution text instruction to speech due to the limitations of the training data distribution in TTS models. To address this issue, we propose a query rewriting framework with multi-LLM knowledge fusion, employing multiple agents to annotate and validate the synthesized speech, making it possible to construct high-quality speech instruction datasets without relying on human annotation. Experiments show that this method can transform text instructions into distributions more suitable for TTS models for speech synthesis through zero-shot rewriting, increasing data usability from 72% to 93%. It also demonstrates unique advantages in rewriting tasks that require complex knowledge and context-related abilities.

Video large language models (VideoLLM) excel at video understanding, but face efficiency challenges due to the quadratic complexity of abundant visual tokens. Our systematic analysis of token compression methods for VideoLLMs reveals two critical issues: (i) overlooking distinctive visual signals across frames, leading to information loss; (ii) suffering from implementation constraints, causing incompatibility with modern architectures or efficient operators.To address these challenges, we distill three design principles for VideoLLM token compression and propose a plug-and-play inference acceleration framework “Video Compression Commander” (VidCom²). By quantifying each frame’s uniqueness, VidCom² adaptively adjusts compression intensity across frames, effectively preserving essential information while reducing redundancy in video sequences. Extensive experiments across various VideoLLMs and benchmarks demonstrate the superior performance and efficiency of our VidCom². With only 25% visual tokens, VidCom² achieves 99.6% of the original performance on LLaVA-OV while reducing 70.8% of the LLM generation latency. Notably, our Frame Compression Adjustment strategy is compatible with other token compression methods to further improve their performance. Our code is available at https://github.com/xuyang-liu16/VidCom2.

Efficient Key-Value (KV) cache management is essential for processing long text sequences in large language models (LLMs), where memory constraints often limit performance. Conventional KV eviction strategies, such as top-k selection based on attention scores, depend on static heuristics that fail to capture the evolving implicit dependencies among tokens during inference. To overcome this, we propose GraphKV, a graph-based framework that redefines token selection for KV cache compression. In GraphKV, tokens are modeled as nodes with importance scores, and edges represent their similarity relationships. Through a decay-signal-propagation mechanism, token importance is dynamically updated by propagating information across the graph, enabling adaptive retention of the most contextually significant tokens. GraphKV can be seamlessly utilized in existing KV cache eviction methods such as SnapKV and PyramidKV in a plug-and-play manner. Codes are available in the supplementary materials and will be released on Github.

Fine-tuning pre-trained Large Language Models (LLMs) for specialized tasks incurs substantial computational and data costs. While model merging offers a training-free solution to integrate multiple task-specific models, existing methods suffer from safety-utility conflicts where enhanced general capabilities degrade safety safeguards. We identify two root causes: neuron misidentification due to simplistic parameter magnitude-based selection, and cross-task neuron interference during merging.To address these challenges, we propose LED-Merging, a three-stage framework that Locates task-specific neurons via gradient-based attribution, dynamically Elects critical neurons through multi-model importance fusion, and Disjoints conflicting updates through parameter isolation.Extensive experiments on Llama-3-8B, Mistral-7B, and Llama2-13B demonstrate that LED-Merging effectively reduces harmful response rates, showing a 31.4% decrease on Llama-3-8B-Instruct on HarmBench, while simultaneously preserving 95% of utility performance, such as achieving 52.39% accuracy on GSM8K.LED-Merging resolves safety-utility conflicts and provides a lightweight, training-free paradigm for constructing reliable multi-task LLMs.Code is available at https://github.com/MqLeet/LED-Merging

Fine-tuning large language models (LLMs) on task-specific data is essential for their effective deployment. As dataset sizes grow, efficiently selecting optimal subsets for training becomes crucial to balancing performance and computational costs. Traditional data selection methods often require fine-tuning a scoring model on the target dataset, which is time-consuming and resource-intensive, or rely on heuristics that fail to fully leverage the model’s predictive capabilities. To address these challenges, we propose Data Whisperer, an efficient, training-free, attention-based method that leverages few-shot in-context learning with the model to be fine-tuned. Comprehensive evaluations were conducted on both raw and synthetic datasets across diverse tasks and models. Notably, Data Whisperer achieves superior performance compared to the full GSM8K dataset on the Llama-3-8B-Instruct model, using just 10% of the data, and outperforms existing methods with a 3.1-point improvement and a 7.4× speedup.

Vision tokens in multimodal large language models often dominate huge computational overhead due to their excessive length compared to linguistic modality. Abundant recent methods aim to solve this problem with token pruning, which first defines an importance criterion for tokens and then prunes the unimportant vision tokens during inference. However, in this paper, we show that the importance is not an ideal indicator to decide whether a token should be pruned. Surprisingly, it usually results in inferior performance than random token pruning and leading to incompatibility to efficient attention computation operators. Instead, we propose DART (Duplication-Aware Reduction of Tokens), which prunes tokens based on its duplication with other tokens, leading to significant and training-free acceleration. Concretely, DART selects a small subset of pivot tokens and then retains the tokens with low duplication to the pivots, ensuring minimal information loss during token pruning. Experiments demonstrate that DART can prune 88.9% vision tokens while maintaining comparable performance, leading to a 1.99× and 2.99× speed-up in total time and prefilling stage, respectively, with good compatibility to efficient attention operators.

Multimodal large language models (MLLMs) have shown remarkable performance for cross-modal understanding and generation, yet still suffer from severe inference costs. Recently, abundant works have been proposed to solve this problem with token pruning, which identifies the redundant tokens in MLLMs and then prunes them to reduce the computation and KV storage costs, leading to significant acceleration without training. While these methods claim efficiency gains, critical questions about their fundamental design and evaluation remain unanswered: Why do many existing approaches underperform even compared to naive random token selection? Are attention-based scoring sufficient for reliably identifying redundant tokens? Is language information really helpful during token pruning? What makes a good trade-off between token importance and duplication? Are current evaluation protocols comprehensive and unbiased? The ignorance of previous research on these problems hinders the long-term development of token pruning. In this paper, we answer these questions one by one, providing insights into the design of future token pruning methods. Codes are available in the supplementary materials.