arXiv subjects

Wentao Bao

Publications and source records attributed to Wentao Bao.

At least 19 recordsLinked to original sources

arXiv2026

MoEMB: Scaling Universal Multimodal Embeddings with Efficient Mixture-of-Experts Models

Universal multimodal embedding (UME) increasingly demands encoder's capacity for handling a broad range of tasks and modalities with increased complexity. Prior scaling methods either increase the representation size, retrieval effort, or scales the encoder into a heavy multimodal LLM. Recent works, such as Think-Then-Embed (TTE), explore scaling via reasoning tokens. However, embedding models are hard to scale up: increasing parameters directly tradeoffs for the large training batch size that contrastive learning needs, and retrieval has to be served under tight latency. Moreover, UME tasks are diverse in complexity, where scaling up embedders can bring significant redundant computation. In this work, we propose MOEMB, which instead scales UME along the expert axis through mixture-of-experts (MoE), growing encoder capacity while preserving single-vector, non-autoregressive encoding. Through a systematic study of the design space and training recipes for MoE-based UME, MoEMB sets a new state of the art on both MMEB-V2 and MRMR among models trained on public MMEB-family data: with only 3B active parameters, MoEMB surpasses TTE-based methods with >4x active parameters, using significantly less computes. To further improve the scalability and efficiency, we conduct the first comprehensive study of adaptive computation for MoE-based embedding, spanning diverse strategies across training-based and inference-only methods. Together, these results support expert scaling as an effective and efficient direction for UME, with adaptive computation further improving efficiency for MLLM-based embedding models towards large-scale retrieval and recommendation systems.

cs.LG↗

arXiv2026

An Event is Worth One Token: Event Tokenization for Industrial-scale LLM Recommendation

LLM-based recommendation has scaled along model capacity and sequence length, yet each position encodes only text, semantic IDs, or a few categorical features, discarding rich user, item, context, and outcome signals available at each event. Under autoregressive modeling, this yields weak queries at each position and, since each position becomes context for the next, the degradation compounds across the sequence. We propose an event-centric paradigm that represents each interaction by its full temporal snapshot, and identify a new scaling dimension we term snapshot resolution: the amount of information encoded per event. To efficiently scale snapshot resolution, we introduce AMBER (Autoregressive Modeling via Bottlenecked Event Representation), which compresses each temporal snapshot into a compact Event Token, a new LLM input modality. The representation is learned end-to-end, while Event Tokens are pre-computed and cached for serving, decoupling snapshot resolution from real-time serving compute. On industrial-scale ranking and retrieval benchmarks, AMBER advances the compute-quality Pareto frontier relative to alternative recommendation paradigms. At sufficient capacity, a single unified tokenizer even outperforms dedicated per-entity tokenizers, demonstrating positive transfer across structurally different entity types. AMBER's Event Tokens also transfer across model architectures: when integrated into a heavily optimized non-LLM ranker as serving-time historical features, they yield statistically significant improvements. Further scaling Event Tokenizer capacity provides additional improvements.

cs.IR↗

arXiv2026

TTE-Flash: Accelerating Reasoning-based Multimodal Representations via Think-Then-Embed Tokens

Recent research has demonstrated that Universal Multimodal Embedding (UME) benefits significantly from Chain-of-Thought (CoT) reasoning. In this paradigm, a generative model produces explicit reasoning traces for a multimodal query, with the final representation extracted from an embedding token attending to both the query and the reasoning. Despite its effectiveness, the computational overhead of generating explicit CoT traces is often prohibitive. In this work, we propose replacing explicit CoT with latent think tokens, which are interpreted as latent variables that can produce explicit CoT traces as observed variables. By optimizing think tokens using CoT generation loss and subsequent embedding tokens using contrastive loss, we produce high-performance, reasoning-aware representations at a constant inference cost. Our study investigates two key architectural designs: 1) how think and embeddings tokens should be extracted from the same LLM backbone. 2) how the tokens should be trained as two dependent tasks. We introduce TTE-Flash-2B, a reasoning-aware multimodal representation model that outperforms its explicit-CoT counterpart on the MMEB-v2 benchmark, while producing latent think tokens that are interpretable both textually and visually. Furthermore, zero-shot evaluation across 15 video datasets reveals scaling behavior as the number of think tokens increases, and motivating a pilot study of adaptive think budget allocation based on task requirements.

cs.AI↗

arXiv2026

BLENDER: Blended Text Embeddings and Diffusion Residuals for Intra-Class Image Synthesis in Deep Metric Learning

The rise of Deep Generative Models (DGM) has enabled the generation of high-quality synthetic data. When used to augment authentic data in Deep Metric Learning (DML), these synthetic samples enhance intra-class diversity and improve the performance of downstream DML tasks. We introduce BLenDeR, a diffusion sampling method designed to increase intra-class diversity for DML in a controllable way by leveraging set-theory inspired union and intersection operations on denoising residuals. The union operation encourages any attribute present across multiple prompts, while the intersection extracts the common direction through a principal component surrogate. These operations enable controlled synthesis of diverse attribute combinations within each class, addressing key limitations of existing generative approaches. Experiments on standard DML benchmarks demonstrate that BLenDeR consistently outperforms state-of-the-art baselines across multiple datasets and backbones. Specifically, BLenDeR achieves 3.7% increase in Recall@1 on CUB-200 and a 1.8% increase on Cars-196, compared to state-of-the-art baselines under standard experimental settings.

cs.CV↗

arXiv2025

Open Set Face Forgery Detection via Dual-Level Evidence Collection

The surge in face forgeries has increasingly undermined confidence in the authenticity of online content. As generation algorithms rapidly evolve, new fake categories will constantly emerge, severely challenging existing face forgery detection methods. Although face forgery detection has recently improved, current techniques remain largely confined to binary Real-vs-Fake classification or the recognition of known fake categories. Moreover, they fail to identify the emergence of entirely new forgery methods. In this work, we study the Open Set Face Forgery Detection (OSFFD) problem, which requires the detection model to identify novel fake categories. To enhance its real-world applicability, we reformulate the OSFFD problem and address it through uncertainty estimation. Specifically, we propose the Dual-Level Evidential face forgery Detection (DLED) approach, which estimates prediction uncertainty by extracting and integrating category-specific evidence on the spatial and frequency levels. Comprehensive experiments across diverse settings demonstrate that our proposed DLED approach achieves state-of-the-art performance. Notably, it surpasses various existing baseline models by a $20\%$ margin on average when identifying forgeries from novel fake categories. Concurrently, our DLED method yields competitive performance on the standard binary Real-versus-Fake face forgery detection task.

cs.CV↗

arXiv2025

Uni-Hand: Universal Hand Motion Forecasting in Egocentric Views

Forecasting how human hands move in egocentric views is critical for applications like augmented reality and human-robot policy transfer. Recently, several hand trajectory prediction (HTP) methods have been developed to generate future possible hand waypoints, which still suffer from insufficient prediction targets, inherent modality gaps, entangled hand-head motion, and limited validation in downstream tasks. To address these limitations, we present a universal hand motion forecasting framework considering multi-modal input, multi-dimensional and multi-target prediction patterns, and multi-task affordances for downstream applications. We harmonize multiple modalities by vision-language fusion, global context incorporation, and task-aware text embedding injection, to forecast hand waypoints in both 2D and 3D spaces. A novel dual-branch diffusion is proposed to concurrently predict human head and hand movements, capturing their motion synergy in egocentric vision. By introducing target indicators, the prediction model can forecast the specific joint waypoints of the wrist or the fingers, besides the widely studied hand center points. In addition, we enable Uni-Hand to additionally predict hand-object interaction states (contact/separation) to facilitate downstream tasks better. As the first work to incorporate downstream task evaluation in the literature, we build novel benchmarks to assess the real-world applicability of hand motion forecasting algorithms. The experimental results on multiple publicly available datasets and our newly proposed benchmarks demonstrate that Uni-Hand achieves the state-of-the-art performance in multi-dimensional and multi-target hand motion forecasting. Extensive validation in multiple downstream tasks also presents its impressive human-robot policy transfer to enable robotic manipulation, and effective feature enhancement for action anticipation/recognition.

cs.CV↗

arXiv2025

Novel Diffusion Models for Multimodal 3D Hand Trajectory Prediction

Predicting hand motion is critical for understanding human intentions and bridging the action space between human movements and robot manipulations. Existing hand trajectory prediction (HTP) methods forecast the future hand waypoints in 3D space conditioned on past egocentric observations. However, such models are only designed to accommodate 2D egocentric video inputs. There is a lack of awareness of multimodal environmental information from both 2D and 3D observations, hindering the further improvement of 3D HTP performance. In addition, these models overlook the synergy between hand movements and headset camera egomotion, either predicting hand trajectories in isolation or encoding egomotion only from past frames. To address these limitations, we propose novel diffusion models (MMTwin) for multimodal 3D hand trajectory prediction. MMTwin is designed to absorb multimodal information as input encompassing 2D RGB images, 3D point clouds, past hand waypoints, and text prompt. Besides, two latent diffusion models, the egomotion diffusion and the HTP diffusion as twins, are integrated into MMTwin to predict camera egomotion and future hand trajectories concurrently. We propose a novel hybrid Mamba-Transformer module as the denoising model of the HTP diffusion to better fuse multimodal features. The experimental results on three publicly available datasets and our self-recorded data demonstrate that our proposed MMTwin can predict plausible future 3D hand trajectories compared to the state-of-the-art baselines, and generalizes well to unseen environments. The code and pretrained models have been released at https://github.com/IRMVLab/MMTwin.

cs.CV↗

arXiv2025

Window Token Concatenation for Efficient Visual Large Language Models

To effectively reduce the visual tokens in Visual Large Language Models (VLLMs), we propose a novel approach called Window Token Concatenation (WiCo). Specifically, we employ a sliding window to concatenate spatially adjacent visual tokens. However, directly concatenating these tokens may group diverse tokens into one, and thus obscure some fine details. To address this challenge, we propose fine-tuning the last few layers of the vision encoder to adaptively adjust the visual tokens, encouraging that those within the same window exhibit similar features. To further enhance the performance on fine-grained visual understanding tasks, we introduce WiCo+, which decomposes the visual tokens in later layers of the LLM. Such a design enjoys the merits of the large perception field of the LLM for fine-grained visual understanding while keeping a small number of visual tokens for efficient inference. We perform extensive experiments on both coarse- and fine-grained visual understanding tasks based on LLaVA-1.5 and Shikra, showing better performance compared with existing token reduction projectors. The code is available: https://github.com/JackYFL/WiCo.

cs.CV↗

arXiv2025

X-SG$^2$S: Safe and Generalizable Gaussian Splatting with X-dimensional Watermarks

3D Gaussian Splatting (3DGS) has been widely used in 3D reconstruction and 3D generation. However, the rapid adoption of 3D Gaussian Splatting raises growing concerns about information leakage and unauthorized use, urging the exploration of effective watermarking techniques. However, existing methods are limited by low capacity, fragility under geometric perturbations, and the infeasible requirement for costly fine-tuning or pipeline modifications, motivating the need for a generalizable, feed-forward framework capable of robust multi-modal embedding with minimal intrusion. In this paper, we propose a new framework X-SG$^2$S which can simultaneously inject 1D to 3D watermarks for copyright protection, while keeping the high fidelity of original 3DGS scenes. Specifically, we first split the watermarks into message patches. A self-adaptive gate is developed to select the injection positions of the watermark messages. Then, we use an XD (multi-dimensional) injection head to inject multi-modal messages into sorted 3DGS points. To restore watermarking messages, a learnable gate is developed to recognize the watermarked locations, from which our XD-extraction heads are used to restore hidden messages. X-SG$^2$S is the first framework to unify 1D-to-3D watermarking and enable simultaneous multi-modal watermark embedding in 3DGS, achieving this with minimal rendering interference and zero modifications to parameters or pipelines. Extensive experiments demonstrate that X-SG$^2$S effectively preserves consistency between the watermark and the original 3DGS, exhibits robustness against model degradation, and maintains accurate judgment capabilities.

cs.CR↗

arXiv2025

Visual Large Language Models for Generalized and Specialized Applications

Visual-language models (VLM) have emerged as a powerful tool for learning a unified embedding space for vision and language. Inspired by large language models, which have demonstrated strong reasoning and multi-task capabilities, visual large language models (VLLMs) are gaining increasing attention for building general-purpose VLMs. Despite the significant progress made in VLLMs, the related literature remains limited, particularly from a comprehensive application perspective, encompassing generalized and specialized applications across vision (image, video, depth), action, and language modalities. In this survey, we focus on the diverse applications of VLLMs, examining their using scenarios, identifying ethics consideration and challenges, and discussing future directions for their development. By synthesizing these contents, we aim to provide a comprehensive guide that will pave the way for future innovations and broader applications of VLLMs. The paper list repository is available: https://github.com/JackYFL/awesome-VLLMs.

cs.CV↗

arXiv2024

Exploiting VLM Localizability and Semantics for Open Vocabulary Action Detection

Action detection aims to detect (recognize and localize) human actions spatially and temporally in videos. Existing approaches focus on the closed-set setting where an action detector is trained and tested on videos from a fixed set of action categories. However, this constrained setting is not viable in an open world where test videos inevitably come beyond the trained action categories. In this paper, we address the practical yet challenging Open-Vocabulary Action Detection (OVAD) problem. It aims to detect any action in test videos while training a model on a fixed set of action categories. To achieve such an open-vocabulary capability, we propose a novel method OpenMixer that exploits the inherent semantics and localizability of large vision-language models (VLM) within the family of query-based detection transformers (DETR). Specifically, the OpenMixer is developed by spatial and temporal OpenMixer blocks (S-OMB and T-OMB), and a dynamically fused alignment (DFA) module. The three components collectively enjoy the merits of strong generalization from pre-trained VLMs and end-to-end learning from DETR design. Moreover, we established OVAD benchmarks under various settings, and the experimental results show that the OpenMixer performs the best over baselines for detecting seen and unseen actions. We release the codes, models, and dataset splits at https://github.com/Cogito2012/OpenMixer.

cs.CV↗

arXiv2024

Learning to Localize Actions in Instructional Videos with LLM-Based Multi-Pathway Text-Video Alignment

Learning to localize temporal boundaries of procedure steps in instructional videos is challenging due to the limited availability of annotated large-scale training videos. Recent works focus on learning the cross-modal alignment between video segments and ASR-transcripted narration texts through contrastive learning. However, these methods fail to account for the alignment noise, i.e., irrelevant narrations to the instructional task in videos and unreliable timestamps in narrations. To address these challenges, this work proposes a novel training framework. Motivated by the strong capabilities of Large Language Models (LLMs) in procedure understanding and text summarization, we first apply an LLM to filter out task-irrelevant information and summarize task-related procedure steps (LLM-steps) from narrations. To further generate reliable pseudo-matching between the LLM-steps and the video for training, we propose the Multi-Pathway Text-Video Alignment (MPTVA) strategy. The key idea is to measure alignment between LLM-steps and videos via multiple pathways, including: (1) step-narration-video alignment using narration timestamps, (2) direct step-to-video alignment based on their long-term semantic similarity, and (3) direct step-to-video alignment focusing on short-term fine-grained semantic similarity learned from general video domains. The results from different pathways are fused to generate reliable pseudo step-video matching. We conducted extensive experiments across various tasks and problem settings to evaluate our proposed method. Our approach surpasses state-of-the-art methods in three downstream tasks: procedure step grounding, step localization, and narration grounding by 5.9\%, 3.1\%, and 2.8\%.

cs.CV↗

arXiv2024

MADiff: Motion-Aware Mamba Diffusion Models for Hand Trajectory Prediction on Egocentric Videos

Understanding human intentions and actions through egocentric videos is important on the path to embodied artificial intelligence. As a branch of egocentric vision techniques, hand trajectory prediction plays a vital role in comprehending human motion patterns, benefiting downstream tasks in extended reality and robot manipulation. However, capturing high-level human intentions consistent with reasonable temporal causality is challenging when only egocentric videos are available. This difficulty is exacerbated under camera egomotion interference and the absence of affordance labels to explicitly guide the optimization of hand waypoint distribution. In this work, we propose a novel hand trajectory prediction method dubbed MADiff, which forecasts future hand waypoints with diffusion models. The devised denoising operation in the latent space is achieved by our proposed motion-aware Mamba, where the camera wearer's egomotion is integrated to achieve motion-driven selective scan (MDSS). To discern the relationship between hands and scenarios without explicit affordance supervision, we leverage a foundation model that fuses visual and language features to capture high-level semantics from video clips. Comprehensive experiments conducted on five public datasets with the existing and our proposed new evaluation metrics demonstrate that MADiff predicts comparably reasonable hand trajectories compared to the state-of-the-art baselines, and achieves real-time performance. We will release our code and pretrained models of MADiff at the project page: https://irmvlab.github.io/madiff.github.io.

cs.CV↗

arXiv2024

Facial Affective Behavior Analysis with Instruction Tuning

Facial affective behavior analysis (FABA) is crucial for understanding human mental states from images. However, traditional approaches primarily deploy models to discriminate among discrete emotion categories, and lack the fine granularity and reasoning capability for complex facial behaviors. The advent of Multi-modal Large Language Models (MLLMs) has been proven successful in general visual understanding tasks. However, directly harnessing MLLMs for FABA is challenging due to the scarcity of datasets and benchmarks, neglecting facial prior knowledge, and low training efficiency. To address these challenges, we introduce (i) an instruction-following dataset for two FABA tasks, e.g., emotion and action unit recognition, (ii) a benchmark FABA-Bench with a new metric considering both recognition and generation ability, and (iii) a new MLLM "EmoLA" as a strong baseline to the community. Our initiative on the dataset and benchmarks reveal the nature and rationale of facial affective behaviors, i.e., fine-grained facial movement, interpretability, and reasoning. Moreover, to build an effective and efficient FABA MLLM, we introduce a facial prior expert module with face structure knowledge and a low-rank adaptation module into pre-trained MLLM. We conduct extensive experiments on FABA-Bench and four commonly-used FABA datasets. The results demonstrate that the proposed facial prior expert can boost the performance and EmoLA achieves the best results on our FABA-Bench. On commonly-used FABA datasets, EmoLA is competitive rivaling task-specific state-of-the-art models.

cs.CV↗

arXiv2023

Latent Space Energy-based Model for Fine-grained Open Set Recognition

Fine-grained open-set recognition (FineOSR) aims to recognize images belonging to classes with subtle appearance differences while rejecting images of unknown classes. A recent trend in OSR shows the benefit of generative models to discriminative unknown detection. As a type of generative model, energy-based models (EBM) are the potential for hybrid modeling of generative and discriminative tasks. However, most existing EBMs suffer from density estimation in high-dimensional space, which is critical to recognizing images from fine-grained classes. In this paper, we explore the low-dimensional latent space with energy-based prior distribution for OSR in a fine-grained visual world. Specifically, based on the latent space EBM, we propose an attribute-aware information bottleneck (AIB), a residual attribute feature aggregation (RAFA) module, and an uncertainty-based virtual outlier synthesis (UVOS) module to improve the expressivity, granularity, and density of the samples in fine-grained classes, respectively. Our method is flexible to take advantage of recent vision transformers for powerful visual classification and generation. The method is validated on both fine-grained and general visual classification datasets while preserving the capability of generating photo-realistic fake images with high resolution.

cs.CV↗

arXiv2023

On Model Explanations with Transferable Neural Pathways

Neural pathways as model explanations consist of a sparse set of neurons that provide the same level of prediction performance as the whole model. Existing methods primarily focus on accuracy and sparsity but the generated pathways may offer limited interpretability thus fall short in explaining the model behavior. In this paper, we suggest two interpretability criteria of neural pathways: (i) same-class neural pathways should primarily consist of class-relevant neurons; (ii) each instance's neural pathway sparsity should be optimally determined. To this end, we propose a Generative Class-relevant Neural Pathway (GEN-CNP) model that learns to predict the neural pathways from the target model's feature maps. We propose to learn class-relevant information from features of deep and shallow layers such that same-class neural pathways exhibit high similarity. We further impose a faithfulness criterion for GEN-CNP to generate pathways with instance-specific sparsity. We propose to transfer the class-relevant neural pathways to explain samples of the same class and show experimentally and qualitatively their faithfulness and interpretability.

cs.CV↗

arXiv2023

Uncertainty-aware State Space Transformer for Egocentric 3D Hand Trajectory Forecasting

Hand trajectory forecasting from egocentric views is crucial for enabling a prompt understanding of human intentions when interacting with AR/VR systems. However, existing methods handle this problem in a 2D image space which is inadequate for 3D real-world applications. In this paper, we set up an egocentric 3D hand trajectory forecasting task that aims to predict hand trajectories in a 3D space from early observed RGB videos in a first-person view. To fulfill this goal, we propose an uncertainty-aware state space Transformer (USST) that takes the merits of the attention mechanism and aleatoric uncertainty within the framework of the classical state-space model. The model can be further enhanced by the velocity constraint and visual prompt tuning (VPT) on large vision transformers. Moreover, we develop an annotation workflow to collect 3D hand trajectories with high quality. Experimental results on H2O and EgoPAT3D datasets demonstrate the superiority of USST for both 2D and 3D trajectory forecasting. The code and datasets are publicly released: https://actionlab-cv.github.io/EgoHandTrajPred.

cs.CV↗

arXiv2023

Prompting Language-Informed Distribution for Compositional Zero-Shot Learning

Compositional zero-shot learning (CZSL) task aims to recognize unseen compositional visual concepts, e.g., sliced tomatoes, where the model is learned only from the seen compositions, e.g., sliced potatoes and red tomatoes. Thanks to the prompt tuning on large pre-trained visual language models such as CLIP, recent literature shows impressively better CZSL performance than traditional vision-based methods. However, the key aspects that impact the generalization to unseen compositions, including the diversity and informativeness of class context, and the entanglement between visual primitives, i.e., state and object, are not properly addressed in existing CLIP-based CZSL literature. In this paper, we propose a model by prompting the language-informed distribution, aka., PLID, for the CZSL task. Specifically, the PLID leverages pre-trained large language models (LLM) to (i) formulate the language-informed class distributions which are diverse and informative, and (ii) enhance the compositionality of the class embedding. Moreover, a visual-language primitive decomposition (VLPD) module is proposed to dynamically fuse the classification decisions from the compositional and the primitive space. Orthogonal to the existing literature of soft, hard, or distributional prompts, our method advocates prompting the LLM-supported class distributions, leading to a better zero-shot generalization. Experimental results on MIT-States, UT-Zappos, and C-GQA datasets show the superior performance of the PLID to the prior arts. Our code and models are released: https://github.com/Cogito2012/PLID.

cs.CV↗