Searcharxiv⌕ Search

arXiv subjects

Mingbao Lin

Publications and source records attributed to Mingbao Lin.

At least 19 recordsLinked to original sources
arXiv2026

Accelerating Unified Multimodal Models with Core-Expansion Routing and Unified Computation Scheduling

Unified multimodal models jointly support understanding and generation, but incur substantial redundant computation across tokens, layers, and generation timesteps. Through token-importance probing, we identify an asymmetric core-expansion structure: understanding exhibits a stable importance component, while generation largely shares this component but requires progress-dependent corrections. We therefore propose CE-Router, which uses a task-shared core scorer and progress-conditioned generation expansions, optimized through generation decomposition and cross-task core alignment. At inference, CE-Router compacts token computation and supplies a learned routing signal to Unified Computation Scheduling, which coordinates layer skipping, FFN pruning, diffusion-head cache reuse, and denoising-step early exit. Experiments on two representative UMM architectures demonstrate consistent quality--efficiency improvements across both tasks, retaining 98.03\% of dense understanding performance with a 1.93$\times$ end-to-end inference speedup.

cs.AI↗
arXiv2026

VoiceMem: Streaming Dual-Brain Memory for Real-Time Interaction

Conversational systems, such as duplex speech language models (SLMs), still lack a streaming, accurate, and empathetic memory system as their soul. We introduce VoiceMem, a simple memory architecture with a parallel informational left brain, an emotional right brain, and streaming memory I/O mechanisms. We further build a complete pipeline for memory-aware SLM training, long-horizon evaluation, and decoupled deployment with interchangeable memory backends. Experiments and real-world deployment show three advantages: i) Accuracy: under top-5 retrieval, the left brain outperforms classical systems such as Mem0 at top-200 by nearly 30 points; ii) Emotional & Personal: the right brain, with short- and long-horizon affective attribution and dual-node persona modeling, achieves state-of-the-art performance across three persona benchmarks and improves the aggregate score by 4.29 points over the previous best system; and iii) Real-Time & Cheap: VoiceMem completes retrieval in 134 ms, well within standard VAD latency, adding no extra conversational delay while maintaining high accuracy and low cost. These results show that VoiceMem provides a practical memory foundation for real-time, personalized, and emotionally aware speech interaction.

eess.AS↗
arXiv2026

HACK++: Towards More Effective Head-Aware Key-Value Compression for Efficient Visual Autoregressive Modeling

Visual Autoregressive (VAR) models adopt a next-scale prediction paradigm, offering high-quality generation with substantially fewer decoding steps. However, existing VAR models suffer from significant attention complexity and severe memory overhead due to the accumulation of key-value (KV) caches across scales. In this paper, we tackle this challenge by introducing KV cache compression into the next-scale paradigm. We begin with an in-depth analysis of VAR attention and observe that attention heads can be stably divided into two functionally distinct categories: Contextual Heads focus on maintaining semantic consistency, while Structural Heads preserve spatial coherence. Their functional divergence makes existing one-size-fits-all compression methods perform poorly on VAR models. We further find that the two head types differ markedly in their reliance on historical scales, and that this reliance shifts across layers and generation steps, arguing for an adaptive cache budget allocation. To address these challenges, we propose HACK++, a training-free Head-Aware key-value Compression frameworK for VAR models. From a one-time offline calibration, HACK++ classifies head types and derives head-specific priors. At inference, it decouples attention from cache compression under independent budgets, bounding the current-scale attention cost while compressing the accumulated cache far more aggressively, via pattern-specific strategies and a reliance-aware budget allocation. Extensive experiments on multiple VAR models across text-to-image, class-conditional, and unified understanding-and-generation tasks validate the effectiveness and generalizability of HACK++. For example, on Infinity-2B/8B, HACK++ maintains near-lossless generation with only a 30% attention budget and a 10% cache budget, and remains robust even under a 1% cache budget.

cs.CV↗
arXiv2026

Audio Interaction Model

Audio is continuous and interactive, yet most Large Audio Language Models (LALMs) remain offline and streaming systems usually specialize in ASR or spoken dialogue. We formalize the Audio Interaction Model, an always-on perceive--decide--respond paradigm that tracks context, decides whether intervention is warranted, and responds without stopping listening. We instantiate it with Audio-Interaction and introduce SoundFlow, coupling streaming-native data construction, comprehension-aware silence/response supervision, dual-loss training, and asynchronous FIFO inference. We also construct textsc{StreamAudio-2M, a 2.6M-item, 302k-hour corpus spanning 7 capability families and 28 sub-tasks, together with Proactive-Sound-Bench. Across 8 benchmarks, Audio-Interaction remains competitive on mainstream audio tasks while enabling spoken-instruction robustness, long-stream interaction, and proactive intervention.

cs.SD↗
arXiv2026

Step-GRPO: Internalizing Dynamic Early Exit for Efficient Reasoning

Large reasoning models that use long chain-of-thought excel at problem-solving yet waste compute on redundant checks. Curbing this overthinking is hard: training-time length penalties can cripple ability, while inference-time early-exit adds system overhead. To bridge this gap, we propose Step-GRPO, a novel post-training framework that internalizes dynamic early-exit capabilities directly into the model. Step-GRPO shifts the optimization objective from raw tokens to semantic steps by utilizing linguistic markers to structure reasoning. We introduce a Dynamic Truncated Rollout mechanism that exposes the model to concise high-confidence trajectories during exploration, synergized with a Step-Aware Relative Reward that dynamically penalizes redundancy based on group-level baselines. Extensive experiments across three model sizes on diverse benchmarks demonstrate that Step-GRPO achieves a superior accuracy-efficiency trade-off. On Qwen3-8B, our method reduces token consumption by 32.0\% compared to the vanilla model while avoiding the accuracy degradation observed in traditional length-penalty methods.

cs.AI↗
arXiv2026

Unbiased Dynamic Pruning for Efficient Group-Based Policy Optimization

Group Relative Policy Optimization (GRPO) effectively scales LLM reasoning but incurs prohibitive computational costs due to its extensive group-based sampling requirement. While recent selective data utilization methods can mitigate this overhead, they could induce estimation bias by altering the underlying sampling distribution, compromising theoretical rigor and convergence behavior. To address this limitation, we propose Dynamic Pruning Policy Optimization (DPPO), a framework that enables dynamic pruning while preserving unbiased gradient estimation through importance sampling-based correction. By incorporating mathematically derived rescaling factors, DPPO significantly accelerates GRPO training without altering the optimization objective of the full-batch baseline. Furthermore, to mitigate the data sparsity induced by pruning, we introduce Dense Prompt Packing, a window-based greedy strategy that maximizes valid token density and hardware utilization. Extensive experiments demonstrate that DPPO consistently accelerates training across diverse models and benchmarks. For instance, on Qwen3-4B trained on MATH, DPPO achieves 2.37$\times$ training speedup and outperforms GRPO by 3.36% in average accuracy across six mathematical reasoning benchmarks.

cs.LG↗
arXiv2026

Reversible Diffusion Decoding for Diffusion Language Models

Diffusion language models enable parallel token generation through block-wise decoding, but their irreversible commitments can lead to stagnation, where the reverse diffusion process fails to make further progress under a suboptimal context.We propose Reversible Diffusion Decoding (RDD), a decoding framework that introduces reversibility into block-wise diffusion generation. RDD detects stagnation as a state-dependent failure of the reverse process and enables efficient backtracking to earlier blocks without recomputation via cached model states. To avoid repeated failure trajectories, RDD applies confidence-guided re-masking to selectively reinitialize uncertain tokens while preserving reliable context.This reversible formulation allows decoding to recover from early commitment errors while maintaining the parallel efficiency of diffusion-based generation. Experiments show that RDD improves generation robustness and quality over baselines with minimal computational overhead.

cs.CL↗
arXiv2025

Autoregressive Image Generation Needs Only a Few Lines of Cached Tokens

Autoregressive (AR) visual generation has emerged as a powerful paradigm for image and multimodal synthesis, owing to its scalability and generality. However, existing AR image generation suffers from severe memory bottlenecks due to the need to cache all previously generated visual tokens during decoding, leading to both high storage requirements and low throughput. In this paper, we introduce \textbf{LineAR}, a novel, training-free progressive key-value (KV) cache compression pipeline for autoregressive image generation. By fully exploiting the intrinsic characteristics of visual attention, LineAR manages the cache at the line level using a 2D view, preserving the visual dependency regions while progressively evicting less-informative tokens that are harmless for subsequent line generation, guided by inter-line attention. LineAR enables efficient autoregressive (AR) image generation by utilizing only a few lines of cache, achieving both memory savings and throughput speedup, while maintaining or even improving generation quality. Extensive experiments across six autoregressive image generation models, including class-conditional and text-to-image generation, validate its effectiveness and generality. LineAR improves ImageNet FID from 2.77 to 2.68 and COCO FID from 23.85 to 22.86 on LlamaGen-XL and Janus-Pro-1B, while retaining only 1/6 KV cache. It also improves DPG on Lumina-mGPT-768 with just 1/8 KV cache. Additionally, LineAR achieves significant memory and throughput gains, including up to 67.61% memory reduction and 7.57x speedup on LlamaGen-XL, and 39.66% memory reduction and 5.62x speedup on Janus-Pro-7B.

cs.CV↗
arXiv2025

Parallel Vision Token Scheduling for Fast and Accurate Multimodal LMMs Inference

Multimodal large language models (MLLMs) deliver impressive vision-language reasoning but suffer steep inference latency because self-attention scales quadratically with sequence length and thousands of visual tokens contributed by high-resolution images. Naively pruning less-informative visual tokens reduces this burden, yet indiscriminate removal can strip away contextual cues essential for background or fine-grained questions, undermining accuracy. In this paper, we present ParVTS (Parallel Vision Token Scheduling), a training-free scheduling framework that partitions visual tokens into subject and non-subject groups, processes them in parallel to transfer their semantics into question tokens, and discards the non-subject path mid-inference to reduce computation. This scheduling reduces computational complexity, requires no heuristics or additional modules, and is compatible with diverse existing MLLM architectures. Experiments across multiple MLLM backbones show that ParVTS prunes up to 88.9% of visual tokens with minimal performance drop, achieving 1.77x speedup and 70% FLOPs reduction.

cs.CV↗
arXiv2025

Test-Time Temporal Sampling for Efficient MLLM Video Understanding

Processing long videos with multimodal large language models (MLLMs) poses a significant computational challenge, as the model's self-attention mechanism scales quadratically with the number of video tokens, resulting in high computational demand and slow inference speed. Current solutions, such as rule-based sub-sampling, learned frame selector, or memory-based summarization, often introduce their own trade-offs: they compromise accuracy, necessitate additional training, or decrease inference speed. In this paper, we propose Test-Time Temporal Sampling (T3S), a training-free, plug-and-play inference wrapper that enables MLLMs to process long videos both efficiently and effectively. T3S exploits spatiotemporal redundancy by generating multiple short and diverse subsequences of video tokens at inference time, packing them within a single forward pass, and aggregating their predictions. This multi-subsequence formulation broadens visual coverage while reducing the computational cost of self-attention from $O(L^2)$ to $O(\sum_{i=1}^m \alpha_i^2L^2)$, where $\sum_{i=1}^m \alpha_i^2 < 1$. Extensive experiments on long video understanding benchmarks demonstrate that T3S improves accuracy by up to 3.1% and reduces first token delay by $2.04\times$, all with minimal integration effort. Our approach operates entirely at inference time, requires no model modifications or fine-tuning, and is compatible with a wide range of pretrained MLLMs. T3S turns video redundancy into a computational advantage, offering a scalable solution for long-video understanding. The code is available at https://github.com/kaibinwang3/T3S.

cs.CV↗
arXiv2025

Progressive Supernet Training for Efficient Visual Autoregressive Modeling

Visual Auto-Regressive (VAR) models significantly reduce inference steps through the "next-scale" prediction paradigm. However, progressive multi-scale generation incurs substantial memory overhead due to cumulative KV caching, limiting practical deployment. We observe a scale-depth asymmetric dependency in VAR: early scales exhibit extreme sensitivity to network depth, while later scales remain robust to depth reduction. Inspired by this, we propose VARiant: by equidistant sampling, we select multiple subnets ranging from 16 to 2 layers from the original 30-layer VAR-d30 network. Early scales are processed by the full network, while later scales utilize subnet. Subnet and the full network share weights, enabling flexible depth adjustment within a single model. However, weight sharing between subnet and the entire network can lead to optimization conflicts. To address this, we propose a progressive training strategy that breaks through the Pareto frontier of generation quality for both subnets and the full network under fixed-ratio training, achieving joint optimality. Experiments on ImageNet demonstrate that, compared to the pretrained VAR-d30 (FID 1.95), VARiant-d16 and VARiant-d8 achieve nearly equivalent quality (FID 2.05/2.12) while reducing memory consumption by 40-65%. VARiant-d2 achieves 3.5 times speedup and 80% memory reduction at moderate quality cost (FID 2.97). In terms of deployment, VARiant's single-model architecture supports zero-cost runtime depth switching and provides flexible deployment options from high quality to extreme efficiency, catering to diverse application scenarios.

cs.CV↗
arXiv2025

Beyond Benchmarks: Understanding Mixture-of-Experts Models through Internal Mechanisms

Mixture-of-Experts (MoE) architectures have emerged as a promising direction, offering efficiency and scalability by activating only a subset of parameters during inference. However, current research remains largely performance-centric, with limited understanding of its internal mechanisms, thereby constraining broader progress. In this work, we use an internal metric to investigate the mechanisms of MoE architecture by explicitly incorporating routing mechanisms and analyzing expert-level behaviors. Through systematic analyses of a wide range of publicly available MoE models, we uncover several findings: (1) neuron utilization decreases as models evolve, reflecting stronger generalization; (2) training exhibits a dynamic trajectory, where benchmark performance alone provides limited signal while MUI reveals deeper insights; (3) task completion emerges from collaborative contributions of multiple experts, with shared experts driving concentration; and (4) activation patterns at the neuron level provide a fine-grained proxy for data diversity. Together, these results demonstrate the potential of MUI as a complementary indicator to benchmark performance, offering new insights into the capacity, dynamics, and specialization of MoE models. Our project can be found at https://yingjiahao14.github.io/MoE-MUI/.

cs.LG↗
arXiv2025

ERTACache: Error Rectification and Timesteps Adjustment for Efficient Diffusion

Diffusion models suffer from substantial computational overhead due to their inherently iterative inference process. While feature caching offers a promising acceleration strategy by reusing intermediate outputs across timesteps, naive reuse often incurs noticeable quality degradation. In this work, we formally analyze the cumulative error introduced by caching and decompose it into two principal components: feature shift error, caused by inaccuracies in cached outputs, and step amplification error, which arises from error propagation under fixed timestep schedules. To address these issues, we propose ERTACache, a principled caching framework that jointly rectifies both error types. Our method employs an offline residual profiling stage to identify reusable steps, dynamically adjusts integration intervals via a trajectory-aware correction coefficient, and analytically approximates cache-induced errors through a closed-form residual linearization model. Together, these components enable accurate and efficient sampling under aggressive cache reuse. Extensive experiments across standard image and video generation benchmarks show that ERTACache achieves up to 2x inference speedup while consistently preserving or even improving visual quality. Notably, on the state-of-the-art Wan2.1 video diffusion model, ERTACache delivers 2x acceleration with minimal VBench degradation, effectively maintaining baseline fidelity while significantly improving efficiency. The code is available at https://github.com/bytedance/ERTACache.

cs.CV↗
arXiv2025

SynergyAmodal: Deocclude Anything with Text Control

Image deocclusion (or amodal completion) aims to recover the invisible regions (\ie, shape and appearance) of occluded instances in images. Despite recent advances, the scarcity of high-quality data that balances diversity, plausibility, and fidelity remains a major obstacle. To address this challenge, we identify three critical elements: leveraging in-the-wild image data for diversity, incorporating human expertise for plausibility, and utilizing generative priors for fidelity. We propose SynergyAmodal, a novel framework for co-synthesizing in-the-wild amodal datasets with comprehensive shape and appearance annotations, which integrates these elements through a tripartite data-human-model collaboration. First, we design an occlusion-grounded self-supervised learning algorithm to harness the diversity of in-the-wild image data, fine-tuning an inpainting diffusion model into a partial completion diffusion model. Second, we establish a co-synthesis pipeline to iteratively filter, refine, select, and annotate the initial deocclusion results of the partial completion diffusion model, ensuring plausibility and fidelity through human expert guidance and prior model constraints. This pipeline generates a high-quality paired amodal dataset with extensive category and scale diversity, comprising approximately 16K pairs. Finally, we train a full completion diffusion model on the synthesized dataset, incorporating text prompts as conditioning signals. Extensive experiments demonstrate the effectiveness of our framework in achieving zero-shot generalization and textual controllability. Our code, dataset, and models will be made publicly available at https://github.com/imlixinyang/SynergyAmodal.

cs.CV↗
arXiv2025

Head-Aware KV Cache Compression for Efficient Visual Autoregressive Modeling

Visual Autoregressive (VAR) models adopt a next-scale prediction paradigm, offering high-quality content generation with substantially fewer decoding steps. However, existing VAR models suffer from significant attention complexity and severe memory overhead due to the accumulation of key-value (KV) caches across scales. In this paper, we tackle this challenge by introducing KV cache compression into the next-scale generation paradigm. We begin with a crucial observation: attention heads in VAR models can be divided into two functionally distinct categories: Contextual Heads focus on maintaining semantic consistency, while Structural Heads are responsible for preserving spatial coherence. This structural divergence causes existing one-size-fits-all compression methods to perform poorly on VAR models. To address this, we propose HACK, a training-free Head-Aware KV cache Compression frameworK. HACK utilizes an offline classification scheme to separate head types, enabling it to apply pattern-specific compression strategies with asymmetric cache budgets for each category. By doing so, HACK effectively constrains the average KV cache length within a fixed budget $B$, reducing the theoretical attention complexity from $\mathcal{O}(n^4)$ to $\mathcal{O}(Bn^2)$. Extensive experiments on multiple VAR models across text-to-image and class-conditional tasks validate the effectiveness and generalizability of HACK. It achieves up to 70% KV cache compression without degrading output quality, resulting in memory savings and faster inference. For example, HACK provides a $1.75\times$ memory reduction and a $1.57\times$ speedup on Infinity-8B.

cs.CV↗
arXiv2025

CPPO: Accelerating the Training of Group Relative Policy Optimization-Based Reasoning Models

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 to sample multiple completions for each question. Our experiment and theoretical analysis reveal 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. Experiments show that CPPO achieves up to $7.98\times$ speedup on GSM8K and $3.48\times$ on Math while preserving or even enhancing the accuracy compared to the original GRPO. We release our code at \href{https://github.com/lzhxmu/CPPO}{https://github.com/lzhxmu/CPPO}.

cs.AI↗
arXiv2025

WideRange4D: Enabling High-Quality 4D Reconstruction with Wide-Range Movements and Scenes

With the rapid development of 3D reconstruction technology, research in 4D reconstruction is also advancing, existing 4D reconstruction methods can generate high-quality 4D scenes. However, due to the challenges in acquiring multi-view video data, the current 4D reconstruction benchmarks mainly display actions performed in place, such as dancing, within limited scenarios. In practical scenarios, many scenes involve wide-range spatial movements, highlighting the limitations of existing 4D reconstruction datasets. Additionally, existing 4D reconstruction methods rely on deformation fields to estimate the dynamics of 3D objects, but deformation fields struggle with wide-range spatial movements, which limits the ability to achieve high-quality 4D scene reconstruction with wide-range spatial movements. In this paper, we focus on 4D scene reconstruction with significant object spatial movements and propose a novel 4D reconstruction benchmark, WideRange4D. This benchmark includes rich 4D scene data with large spatial variations, allowing for a more comprehensive evaluation of the generation capabilities of 4D generation methods. Furthermore, we introduce a new 4D reconstruction method, Progress4D, which generates stable and high-quality 4D results across various complex 4D scene reconstruction tasks. We conduct both quantitative and qualitative comparison experiments on WideRange4D, showing that our Progress4D outperforms existing state-of-the-art 4D reconstruction methods. Project: https://github.com/Gen-Verse/WideRange4D

cs.CV↗
arXiv2025

CAKE: Cascading and Adaptive KV Cache Eviction with Layer Preferences

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 10x speedup in decoding latency compared to full cache when processing contexts of 128K tokens with FlashAttention-2. Our code is available at https://github.com/antgroup/cakekv.

cs.CL↗