Zeroth-Order optimization presents a promising memory-efficient paradigm for fine-tuning Large Language Models by relying solely on forward passes. However, its practical adoption is severely constrained by slow wall-clock convergence and high estimation variance. In this work, we dissect the runtime characteristics of ZO algorithms and identify a critical system bottleneck where the generation of perturbations and parameter updates accounts for over 40% of the training latency. We argue that the standard uniform exploration strategy is fundamentally flawed as it fails to account for the heterogeneous sensitivity of layers in deep networks, resulting in computationally wasteful blind searches. To address this structural mismatch, we propose AdaLeZO, an Adaptive Layer-wise ZO optimization framework. By formulating the layer selection process as a non-stationary Multi-Armed Bandit problem, AdaLeZO dynamically allocates the limited perturbation budget to the most sensitive parameters. We further introduce an Inverse Probability Weighting mechanism based on sampling with replacement, which guarantees unbiased gradient estimation while effectively acting as a temporal denoiser to reduce variance. Extensive experiments on LLaMA and OPT models ranging from 6.7B to 30B parameters demonstrate that AdaLeZO achieves 1.7x to 3.0x wall-clock acceleration compared to state-of-the-art methods. Crucially, AdaLeZO functions as a universal plug-and-play module that seamlessly enhances the efficiency of existing ZO optimizers without incurring additional memory overhead.

Human-Object Interaction (HOI) video generation aims to synthesize realistic videos of humans manipulating diverse objects, serving as a promising avenue for AI-driven live streaming e-commerce. A primary obstacle in this domain lies in the complexity of modeling fine-grained physical dynamics and the intricate spatial-temporal coordination between human hands and objects. Existing approaches to this problem typically rely on dense temporal guidance, e.g., frame-wise hand-object pose sequences, to strictly control the interaction process. However, such dense guidance incurs high annotation costs and affects motion synthesis diversity. To overcome these limitations, we introduce SparseCtrl-HOI, a novel sparse temporal control framework for HOI video generation. It requires only a few keyframes that capture interaction states at designated timestamps. Specifically, we employ a Time-Controlled Rotary Positional Embedding (TiRoPE) mechanism to temporally anchor these keyframes while preserving their spatial integrity. Subsequently, to govern the dynamics across intermediate frames, we propose a Motion Prior Injection Module that leverages Multimodal Large Language Models (MLLMs) to extract high-level motion priors. This empowers the model to hallucinate logically and physically plausible transitions. Furthermore, we build SparseHOI-5K, a high-quality and richly annotated dataset for HOI video generation with sparse temporal control. Comprehensive evaluations confirm that our method substantially reduces annotation overhead while synthesizing superior live-streaming e-commerce videos. Both our code and dataset are publicly available at https://mpi-lab.github.io/SparseCtrl-HOI.

Generating realistic and physically plausible 3D Human-Object Interactions (HOI) remains a key challenge in motion generation. One primary reason is that describing these physical constraints with words alone is difficult. To address this limitation, we propose a new paradigm: extracting rich interaction priors from easily accessible 2D images. Specifically, we introduce ViHOI, a novel framework that enables diffusion-based generative models to leverage rich, task-specific priors from 2D images to enhance generation quality. We utilize a large Vision-Language Model (VLM) as a powerful prior-extraction engine and adopt a layer-decoupled strategy to obtain visual and textual priors. Concurrently, we design a Q-Former-based adapter that compresses the VLM’s high-dimensional features into compact prior tokens, which significantly facilitates the conditional training of our diffusion model. Our framework is trained on motion-rendered images from the dataset to ensure strict semantic alignment between visual inputs and motion sequences. During inference, it leverages reference images synthesized by a text-to-image generation model to improve generalization to unseen objects and interaction categories. Experimental results demonstrate that ViHOI achieves state-of-the-art performance, outperforming existing methods across multiple benchmarks and demonstrating superior generalization. The code for this work will be released.

Trajectory-controlled human motion generation aims to synthesize realistic human motions conditioned on both textual descriptions and spatial trajectories. However, existing methods suffer from two critical limitations: first, the conflict between text and trajectory conditions disrupts the denoising process, resulting in compromised motion quality or inaccurate trajectory following; second, the use of redundant motion representations introduces inconsistencies between motion components, leading to instability during trajectory control. To address these challenges, we propose CMC, a decoupled framework that effectively coordinates text and trajectory conditions through a divide-and-conquer strategy. CMC follows a divide-and-conquer paradigm, comprising two cascaded stages: Trajectory Control and Motion Completion. In the first stage, a diffusion model generates a simplified representation of the controlled joints under trajectory guidance, based on the given trajectories, ensuring accurate and stable trajectory following. In the second stage, a text-conditioned diffusion inpainting model generates full-body motions using the simplified representation from the first stage as partial observations. To mitigate overfitting caused by limited inpainting training data, we further introduce the Selective Inpainting Mechanism (SIM), which alternates between text-to-motion generation and motion inpainting tasks during training. Experiments on HumanML3D and KIT datasets demonstrate that CMC achieves state-of-the-art performance in control accuracy and motion quality, demonstrating its effectiveness in coordinating multimodal conditions and representations.

Open-vocabulary human-object interaction (HOI) detection aims to localize and recognize all human-object interactions in an image, including those unseen during training.Existing approaches usually rely on the collaboration between a conventional HOI detector and a Vision-Language Model (VLM) to recognize unseen HOI categories.However, feature fusion in this paradigm is challenging due to significant gaps in cross-model representations.To address this issue, we introduce SL-HOI , a S tream L ined open-vocabulary HOI detection framework based solely on the powerful DINOv3 model.Our design leverages the complementary strengths of DINOv3's components: its backbone for fine-grained localization and its text-aligned vision head for open-vocabulary interaction classification.Moreover, to facilitate smooth cross-attention between the interaction queries and the vision head's output, we propose first feeding both the interaction queries and the backbone patch tokens into the vision head, effectively bridging their representation gaps. All DINOv3 parameters in our approach are frozen, with only a small number of learnable parameters added, allowing a fast adaptation to the HOI detection task.Extensive experiments show that SL-HOI achieves state-of-the-art performance on both the SWiG-HOI and HICO-DET benchmarks, demonstrating the effectiveness of our streamlined model architecture.The code of this work will be released soon.

Diffusion models have seen widespread adoption for text-driven human motion generation and related tasks due to their impressive generative capabilities and flexibility. However, current motion diffusion models face two major limitations: a representational gap caused by pre-trained text encoders that lack motion-specific information, and error propagation during the iterative denoising process. This paper introduces Reconstruction-Anchored Diffusion Model (RAM) to address these challenges. First, RAM leverages a motion latent space as intermediate supervision for text-to-motion generation. To this end, RAM co-trains a motion reconstruction branch with two key objective functions: self-regularization to enhance the discrimination of the motion space and motion-centric latent alignment to enable accurate mapping from text to the motion latent space. Second, we propose Reconstructive Error Guidance (REG), a testing-stage guidance mechanism that exploits the diffusion model's inherent self-correction ability to mitigate error propagation. At each denoising step, REG uses the motion reconstruction branch to reconstruct the previous estimate, reproducing the prior error patterns. By amplifying the residual between the current prediction and the reconstructed estimate, REG highlights the improvements in the current prediction. Extensive experiments demonstrate that RAM achieves significant improvements and state-of-the-art performance. Our code will be released.

Human-object interaction (HOI) synthesis is crucial for creating immersive and realistic experiences for applications such as virtual reality. Existing methods often rely on simplified object representations, such as the object's centroid or the nearest point to a human, to achieve physically plausible motions. However, these approaches may overlook geometric complexity, resulting in suboptimal interaction fidelity. To address this limitation, we introduce ROG, a novel diffusion-based framework that models the spatiotemporal relationships inherent in HOIs with rich geometric detail. For efficient object representation, we select boundary-focused and fine-detail key points from the object mesh, ensuring a comprehensive depiction of the object's geometry. This representation is used to construct an interactive distance field (IDF), capturing the robust HOI dynamics. Furthermore, we develop a diffusion-based relation model that integrates spatial and temporal attention mechanisms, enabling a better understanding of intricate HOI relationships. This relation model refines the generated motion's IDF, guiding the motion generation process to produce relation-aware and semantically aligned movements. Experimental evaluations demonstrate that ROG significantly outperforms state-of-the-art methods in the realism and semantic accuracy of synthesized HOIs. This paper's code will be released at https://lalalfhdh.github.io/rog page/.

Human preference alignment can significantly enhance the capabilities of Multimodal Large Language Models (MLLMs). However, collecting high-quality preference data remains costly. One promising solution is the self-evolution strategy, where models are iteratively trained on data they generate. Current multimodal self-evolution techniques, nevertheless, still need human- or GPT-annotated data. Some methods even require extra models or ground truth answers to construct preference data. To overcome these limitations, we propose a novel multimodal self-evolution framework that empowers the model to autonomously generate high-quality questions and answers using only unannotated images. First, in the question generation phase, we implement an image-driven self-questioning mechanism. This approach allows the model to create questions and evaluate their relevance and answerability based on the image content. If a question is deemed irrelevant or unanswerable, the model regenerates it to ensure alignment with the image. This process establishes a solid foundation for subsequent answer generation and optimization. Second, while generating answers, we design an answer self-enhancement technique to boost the discriminative power of answers. We begin by captioning the images and then use the descriptions to enhance the generated answers. Additionally, we utilize corrupted images to generate rejected answers, thereby forming distinct preference pairs for effective optimization. Finally, in the optimization step, we incorporate an image content alignment loss function alongside the Direct Preference Optimization (DPO) loss to mitigate hallucinations. This function maximizes the likelihood of the above generated descriptions in order to constrain the model's attention to the image content. As a result, model can generate more accurate and reliable outputs. Experiments demonstrate that our framework is competitively compared with previous methods that utilize external information, paving the way for more efficient and scalable MLLMs.

Open vocabulary Human-Object Interaction (HOI) detection is a challenging task that detects all triplets of interest in an image, even those that are not pre-defined in the training set. Existing approaches typically rely on output features generated by large Vision-Language Models (VLMs) to enhance the generalization ability of interaction representations. However, the visual features produced by VLMs are holistic and coarse-grained, which contradicts the nature of detection tasks. To address this issue, we propose a novel Bilateral Collaboration framework for open vocabulary HOI detection (BC-HOI). This framework includes an Attention Bias Guidance (ABG) component, which guides the VLM to produce fine-grained instance-level interaction features according to the attention bias provided by the HOI detector. It also includes a Large Language Model (LLM)-based Supervision Guidance (LSG) component, which provides fine-grained token-level supervision for the HOI detector by the LLM component of the VLM. LSG enhances the ability of ABG to generate high-quality attention bias. We conduct extensive experiments on two popular benchmarks: HICO-DET and V-COCO, consistently achieving superior performance in the open vocabulary and closed settings. The code will be released in Github.

Co-speech gesture video generation aims to synthesize realistic, audio-aligned videos of speakers, complete with synchronized facial expressions and body gestures. This task presents challenges due to the significant one-to-many mapping between audio and visual content, further complicated by the scarcity of large-scale public datasets and high computational demands. We propose a lightweight framework that utilizes 2D full-body skeletons as an efficient auxiliary condition to bridge audio signals with visual outputs. Our approach introduces a diffusion model conditioned on fine-grained audio segments and a skeleton extracted from the speaker's reference image, predicting skeletal motions through skeleton-audio feature fusion to ensure strict audio coordination and body shape consistency. The generated skeletons are then fed into an off-the-shelf human video generation model with the speaker's reference image to synthesize high-fidelity videos. To democratize research, we present CSG-405-the first public dataset with 405 hours of high-resolution videos across 71 speech types, annotated with 2D skeletons and diverse speaker demographics. Experiments show that our method exceeds state-of-the-art approaches in visual quality and synchronization while generalizing across speakers and contexts.

Long-term test-time adaptation (TTA) is a challenging task due to error accumulation. Recent approaches tackle this issue by actively labeling a small proportion of samples in each batch, yet the annotation burden quickly grows as the batch number increases. In this paper, we investigate how to achieve effortless active labeling so that a maximum of one sample is selected for annotation in each batch. First, we annotate the most valuable sample in each batch based on the single-step optimization perspective in the TTA context. In this scenario, the samples that border between the source- and target-domain data distributions are considered the most feasible for the model to learn in one iteration. Then, we introduce an efficient strategy to identify these samples using feature perturbation. Second, we discover that the gradient magnitudes produced by the annotated and unannotated samples have significant variations. Therefore, we propose balancing their impact on model optimization using two dynamic weights. Extensive experiments on the popular ImageNet-C, -R, -K, -A and PACS databases demonstrate that our approach consistently outperforms state-of-the-art methods with significantly lower annotation costs. Code is available at: https://github.com/flash1803/EATTA.

Large Language Models (LLMs) excel at natural language processing tasks, but their massive size leads to high computational and storage demands. Recent works have sought to reduce their model size through layer-wise structured pruning. However, they tend to ignore retaining the capabilities in the pruned part. In this work, we re-examine structured pruning paradigms and uncover several key limitations: 1) notable performance degradation due to direct layer removal, 2) incompetent linear weighted layer aggregation, and 3) the lack of effective post-training recovery mechanisms. To address these limitations, we propose CoMe, including a progressive layer pruning framework with a Concatenation-based Merging technology and a hierarchical distillation post-training process. Specifically, we introduce a channel sensitivity metric that utilizes activation intensity and weight norms for fine-grained channel selection. Subsequently, we employ a concatenation-based layer merging method to fuse the most critical channels in the adjacent layers, enabling a progressive model size reduction. Finally, we propose a hierarchical distillation protocol, which leverages the correspondences between the original and pruned model layers established during pruning, enabling efficient knowledge transfer. Experiments on seven benchmarks show that CoMe achieves state-of-the-art performance; when pruning 30% of LLaMA-2-7b's parameters, the pruned model retains 83% of its original average accuracy.

The chart-to-code generation task requires MLLMs to convert chart images into executable code. This task faces two major challenges: limited data diversity and insufficient maintenance of visual consistency between generated and original charts during training. Existing datasets mainly rely on seed data to prompt GPT models for code generation, resulting in homogeneous samples. To address this, we propose ReChartPrompt, which leverages real-world, human-designed charts from arXiv papers as prompts instead of synthetic seeds. Using the diverse styles and rich content of arXiv charts, we construct ReChartPrompt-240K, a large-scale and highly diverse dataset. Another challenge is that although SFT effectively improve code understanding, it often fails to ensure that generated charts are visually consistent with the originals. To address this, we propose ChartSimRL, a GRPO-based reinforcement learning algorithm guided by a novel chart similarity reward. This reward consists of attribute similarity, which measures the overlap of chart attributes such as layout and color between the generated and original charts, and visual similarity, which assesses similarity in texture and other overall visual features using convolutional neural networks. Unlike traditional text-based rewards such as accuracy or format rewards, our reward considers the multimodal nature of the chart-to-code task and effectively enhances the model's ability to accurately reproduce charts. By integrating ReChartPrompt and ChartSimRL, we develop the ChartMaster model, which achieves state-of-the-art results among 7B-parameter models and even rivals GPT-4o on various chart-to-code generation benchmarks. All resources are available at https://github.com/WentaoTan/ChartMaster.

Achieving human-like reasoning capabilities in Multimodal Large Language Models (MLLMs) has long been a goal. Current methodologies primarily focus on synthesizing positive rationales, while overlooking the critical role of negative rationales in training models to discern flawed reasoning patterns. To address this gap, we propose a novel framework: Self-Aligning Multimodal Reasoning with Answer-Oriented Chain-of-Thought (SMART). This framework enables models to utilize AoT-Oriented Chain-of-Thought (AoT) prompts to automatically generate high-quality positive and negative reasoning paths, followed by self-alignment to enhance their reasoning abilities. Inspired by human strategies for solving proof-based problems, AoT uses answers as a guide to help the model extract critical visual information that links questions and answers. When provided with ground truth answers, the model produces strong positive rationales. Conversely, when correct answers are replaced with misleading alternatives, the model generates an erroneous yet compelling reasoning path, serving as a form of discriminative negative rationale. Models trained with AoT-generated data outperform those trained on manually annotated datasets, demonstrating superior reasoning capabilities. This encourages the use of improved models to generate higher-quality preference data for further optimization. Consequently, SMART establishes an iterative generation-optimization method that continually enhances the model's reasoning skills. Experiments indicate that the SMART framework significantly improves various MLLMs, regardless of model architecture, parameter size, or pre-training dataset. The code, datasets, and models will be released.

Test-time adaptation (TTA) is a task that continually adapts a pre-trained source model to the target domain during inference. One popular approach involves fine-tuning model with cross-entropy loss according to estimated pseudo-labels. However, its performance is significantly affected by noisy pseudo-labels. This study reveals that minimizing the classification error of each sample causes the cross-entropy loss's vulnerability to label noise. To address this issue, we propose a novel Decoupled Prototype Learning (DPL) method that features prototype-centric loss computation. First, we decouple the optimization of class prototypes. For each class prototype, we reduce its distance with positive samples and enlarge its distance with negative samples in a contrastive manner. This strategy prevents the model from overfitting to noisy pseudo-labels. Second, we propose a memory-based strategy to enhance DPL's robustness for the small batch sizes often encountered in TTA. We update each class's pseudo-feature from a memory in a momentum manner and insert an additional DPL loss. Finally, we introduce a consistency regularization-based approach to leverage samples with unconfident pseudo-labels. This approach transfers feature styles of samples with unconfident pseudo-labels to those with confident pseudo-labels. Thus, more reliable samples for TTA are created. The experimental results demonstrate that our methods achieve state-of-the-art performance on domain generalization benchmarks, and reliably improve the performance of self-training-based methods on image corruption benchmarks. The code will be released.

Text-to-image person re-identification (ReID) aims to retrieve the images of an interested person based on textual descriptions. One main challenge for this task is the high cost in manually annotating large-scale databases, which affects the generalization ability of ReID models. Recent works handle this problem by leveraging Multi-modal Large Language Models (MLLMs) to describe pedestrian images automatically. However, the captions produced by MLLMs lack diversity in description styles. To address this issue, we propose a Human Annotator Modeling (HAM) approach to enable MLLMs to mimic the description styles of thousands of human annotators. Specifically, we first extract style features from human textual descriptions and perform clustering on them. This allows us to group textual descriptions with similar styles into the same cluster. Then, we employ a prompt to represent each of these clusters and apply prompt learning to mimic the description styles of different human annotators. Furthermore, we define a style feature space and perform uniform sampling in this space to obtain more diverse clustering prototypes, which further enriches the diversity of the MLLM-generated captions. Finally, we adopt HAM to automatically annotate a massive-scale database for text-to-image ReID. Extensive experiments on this database demonstrate that it significantly improves the generalization ability of ReID models. Code is available at https://github.com/sssaury/HAM.

Diffusion-based models, widely used in text-to-image generation, have proven effective in 2D representation learning. Recently, this framework has been extended to 3D self-supervised learning by constructing a conditional point generator for enhancing 3D representations. However, its performance remains constrained by the 3D diffusion model, which is trained on the available 3D datasets with limited size. We hypothesize that the robust capabilities of text-to-image diffusion models, particularly Stable Diffusion (SD), which is trained on large-scale datasets, can help overcome these limitations. To investigate this hypothesis, we propose PointSD, a framework that leverages the SD model for 3D self-supervised learning. By replacing the SD model's text encoder with a 3D encoder, we train a point-to-image diffusion model that allows point clouds to guide the denoising of rendered noisy images. With the trained point-to-image diffusion model, we use noise-free images as the input and point clouds as the condition to extract SD features. Next, we train a 3D backbone by aligning its features with these SD features, thereby facilitating direct semantic learning. Comprehensive experiments on downstream point cloud tasks and ablation studies demonstrate that the SD model can enhance point cloud self-supervised learning. Code is publicly available at https://github.com/wdttt/PointSD.

Existing test-time adaptation (TTA) methods primarily focus on scenarios involving domain shifts in a single modality. However, they often prove ineffective when multiple modalities simultaneously undergo domain shifts, as they struggle to identify and utilize reliable samples within testing batches amid severe prediction bias. To address this problem, we propose Partition-Then-Adapt (PTA), a novel approach combating prediction bias for TTA with multi-modal domain shifts. PTA comprises two key components: Partition and Debiased Reweighting (PDR) and multi-modal Attention-Guided Alignment (AGA). Specifically, PDR evaluates each sample’s predicted label frequency relative to the batch average, partitioning the batch into potential reliable and unreliable subsets. It then reweights each sample by jointly assessing its bias and confidence levels through a quantile-based approach. By applying weighted entropy loss, PTA simultaneously promotes learning from reliable subsets and discourages reliance on unreliable ones. Moreover, AGA regularizes PDR to focus on semantically meaningful multi-modal cues. Extensive experiments validate the effectiveness of PTA, surpassing state-of-the-art method by 6.1\% on Kinetics50-MC and 5.8\% on VGGSound-MC, respectively. Code of this paper is available at https://github.com/MPI-Lab/PTA.

Image-based virtual try-on is an increasingly important task for online shopping. It aims to synthesize images of a specific person wearing a specified garment. Diffusion model-based approaches have recently become popular, as they are excellent at image synthesis tasks. However, these approaches usually employ additional image encoders and rely on the cross-attention mechanism for texture transfer from the garment to the person image, which affects the try-on's efficiency and fidelity. To address these issues, we propose an Texture-Preserving Diffusion (TPD) model for virtual try-on, which enhances the fidelity of the results and introduces no additional image encoders. Accordingly, we make contributions from two aspects. First, we propose to concatenate the masked person and reference garment images along the spatial dimension and utilize the resulting image as the input for the diffusion model's denoising UNet. This enables the original self-attention layers contained in the diffusion model to achieve efficient and accurate texture transfer. Second, we propose a novel diffusion-based method that predicts a precise inpainting mask based on the person and reference garment images, further enhancing the reliability of the try-on results. In addition, we integrate mask prediction and image synthesis into a single compact model. The experimental results show that our approach can be applied to various try-on tasks, e.g., garment-to-person and person-to-person try-ons, and significantly outperforms state-of-the-art methods on popular VITON, VITON-HD databases.

Achieving 3D hand-object pose estimation in interaction scenarios is challenging due to the severe occlusion generated during the interaction. Existing methods address this issue by utilizing the correlation between the hand and object poses as additional cues. They usually first extract the hand and object features from their respective regions and then refine them with each other. However, this paradigm disregards the role of a broad range of image context. To address this problem, we propose a novel and robust approach that learns a broad range of context by imposing priors. First, we build this approach using stacked transformer decoder layers. These layers are required for extracting image-wide context and regional hand or object features by constraining cross-attention operations. We share the context decoder layer parameters between the hand and object pose estimations to avoid interference in the context-learning process. This imposes a prior, indicating that the hand and object are mutually the most important context for each other, significantly enhancing the robustness of obtained context features. Second, since they play different roles, we provide customized feature maps for the context, hand, and object decoder layers. This strategy facilitates the disentanglement of these layers, reducing the feature learning complexity. Finally, we conduct extensive experiments on the popular HO3D and Dex-YCB databases. The experimental results indicate that our method significantly outperforms state-of-the-art approaches and can be applied to other hand pose estimation tasks. Code is available at https://github.com/zskuang58/LCP.