Searcharxiv⌕ Search

arXiv subjects

Jinhui Tang

Publications and source records attributed to Jinhui Tang.

At least 19 recordsLinked to original sources
arXiv2026

LIBERO-VIFO: Benchmarking the Capability and Safety of Visual Cue Following in Vision-Language-Action Models

Visual cues are increasingly adopted to guide robot learning, but whether Vision-Language-Action (VLA) models can reliably follow authorized cues while disregarding unauthorized ones remains unclear. Existing work covers only a narrow range of cue forms and focuses on final task success, providing only a coarse assessment of cue-following capability. Treating all visual cues as authorized also leaves safety risks of unauthorized following unexplored. To address these gaps, we introduce LIBERO-VIFO, a benchmark to evaluate both the capability and safety of visual cue following in VLA models. LIBERO-VIFO defines eight visual cue families spanning diverse forms. A total of four protocols in two parts are defined: Part I tests cue understanding and authorized following, while Part II evaluates unauthorized visual cue following under language-cue conflict and empty language conditions. Evaluating seven VLA models reveals that although visual cue understanding does not reliably translate into execution, current VLAs are able to execute cue-indicated tasks without language instruction, exposing an emerging risk of unauthorized visual cue following. Extended experiments on scene-instantiated cues, safety-critical settings, and real-robot deployment corroborate these findings. LIBERO-VIFO brings both the capability and safety of visual cue following into systematic evaluation, establishing visual-centric safety as a new perspective for the VLA community.

cs.RO↗
arXiv2026

SafeNexus: Discovering and Steering Modality-Universal Safety Neurons in MLLMs

Although Large Language Models (LLMs) have demonstrated promising safety performance, extending them to Multimodal Large Language Models (MLLMs) exposes a significant gap between expanded multimodal capabilities and existing safety mechanisms. Current defenses remain predominantly confined to specific modal settings, thereby limiting their robustness against broader cross-modal threats. To bridge this gap, we introduce SafeNexus, a cross-modal safety alignment framework that adopts a dedicated neuron-level intervention strategy. First, we formulate a neuron localization paradigm that identifies functionally specialized neurons by characterizing intermediate-layer activation patterns and quantifying their functional salience through importance scoring. Building upon this paradigm, we exploit contrastive data to identify modality-bound safety neurons (BS-Neurons), and validate their role in regulating safety behavior within each modality via targeted suppression. Further cross-modal analysis defines modality-universal safety neurons (US-Neurons) as the shared subset of BS-Neurons identified across individual modalities, serving as the core for defending against harmful cross-modal attacks. We observe that suppressing these neurons substantially degrades safety performance across modalities, while leaving overall utility largely unaffected. Building on these insights, we propose two safety alignment strategies: activation-level safety amplifier and safety neuron calibrator. The proposed strategies enhance model safety through two distinct routes: the former amplifies the activation magnitudes of US-Neurons, while the latter selectively calibrates them via targeted fine-tuning. Extensive experiments demonstrate that our method outperforms prevailing state-of-the-art approaches on safety benchmarks spanning diverse modality combinations, while effectively preserving utility.

cs.CV↗
arXiv2026

One Anchor for All: Unified Multilingual and Multimodal Safety Alignment for LVLMs

As large vision-language models (LVLMs) are deployed globally, the combination of multilingual instructions and visual information makes malicious attacks more covert and sophisticated than ever before. However, existing methods isolate language and modality defenses, which, coupled with the scarcity of safety data and high fine-tuning costs, makes it difficult for models to defend against compound attacks. To address this severe challenge, we propose a neuron-level cross-dimensional safety alignment framework driven by modality- and language-shared safety neurons (MLS-Neurons). First, we identify monolingual and unimodal safety neurons by comparing responses to harmful and benign samples, quantifying functional saliency through activation strength and downstream impact. Then, by intersecting these unimodal neurons within each language, we extract modality-shared safety neurons (MS-Neurons) responsive to both visual and textual risks, bridging the safety representation gap between modalities. Furthermore, using English as a semantic anchor, we intersect MS-Neurons across languages to identify modality- and language-shared safety neurons (MLS-Neurons), serving as key defenses against compound attacks. Finally, we update only this minimal subset of shared neurons (~0.03% of parameters), transferring English-only safety supervision to multilingual and multimodal scenarios. Extensive experiments show that our method significantly outperforms state-of-the-art approaches across diverse multilingual and multimodal safety benchmarks while preserving general utility.

cs.AI↗
arXiv2026

Targeted Interpretable Safety Neuron Enhancement for Multilingual Vision-Language Large Models

With the widespread deployment of vision-language large models (VLLMs), their safety alignment faces dual challenges across languages and modalities. Existing methods model multilingual and multimodal safety separately, overlooking coupled risks between low-resource-language instructions and visual contexts, which hinders the detection of cross-lingual and cross-modal harmful intent and the formation of robust safety boundaries. To address this, we propose a neuron-level interpretable safety alignment framework that identifies safety neurons and performs neuron-targeted safety tuning to jointly mitigate multilingual and multimodal risks. Specifically, we compare FFN representations elicited by harmful requests and benign inputs to identify neuron activation strengths associated with safety refusals. Next, we jointly model neuron activations and corresponding down-projection columns to derive neuron-level saliency, separating general multilingual and multimodal neurons from safety neurons responsible for model defense. Finally, neuron-targeted gradient masking restricts parameter updates to the safety subspace spanned by the identified neurons, enabling precise and interpretable safety enhancement. Extensive experiments show that our method enhances multilingual and multimodal safety by tuning only a few safety neurons, while preserving general capabilities.

cs.CV↗
arXiv2026

MAB-DQA: Addressing Query Aspect Importance in Document Question Answering with Multi-Armed Bandits

Document Question Answering (DQA) involves generating answers from a document based on a user's query, representing a key task in document understanding. This task requires interpreting visual layouts, which has prompted recent studies to adopt multimodal Retrieval-Augmented Generation (RAG) that processes page images for answer generation. However, in multimodal RAG, visual DQA struggles to utilize a large number of images effectively, as the retrieval stage often retains only a few candidate pages (e.g., Top-4), causing informative but less visually salient content to be overlooked in favor of common yet low-information pages. To address this issue, we propose a Multi-Armed Bandit-based DQA framework (MAB-DQA) to explicitly model the varying importance of multiple implicit aspects in a query. Specifically, MAB-DQA decomposes a query into aspect-aware subqueries and retrieves an aspect-specific candidate set for each. It treats each subquery as an arm and uses preliminary reasoning results from a small number of representative pages as reward signals to estimate aspect utility. Guided by an exploration-exploitation policy, MAB-DQA dynamically reallocates retrieval budgets toward high-value aspects. With the most informative pages and their correlations, MAB-DQA generates the expected results. On four benchmarks, MAB-DQA shows an average improvement of 5%-18% over the state-of-the-art method, consistently enhancing document understanding. Codes are available at https://github.com/ElephantOH/MAB-DQA.

cs.CL↗
arXiv2026

VersaVogue: Visual Expert Orchestration and Preference Alignment for Unified Fashion Synthesis

Diffusion models have driven remarkable advancements in fashion image generation, yet prior works usually treat garment generation and virtual dressing as separate problems, limiting their flexibility in real-world fashion workflows. Moreover, fashion image synthesis under multi-source heterogeneous conditions remains challenging, as existing methods typically rely on simple feature concatenation or static layer-wise injection, which often causes attribute entanglement and semantic interference. To address these issues, we propose VersaVogue, a unified framework for multi-condition controllable fashion synthesis that jointly supports garment generation and virtual dressing, corresponding to the design and showcase stages of the fashion lifecycle. Specifically, we introduce a trait-routing attention (TA) module that leverages a mixture-of-experts mechanism to dynamically route condition features to the most compatible experts and generative layers, enabling disentangled injection of visual attributes such as texture, shape, and color. To further improve realism and controllability, we develop an automated multi-perspective preference optimization (MPO) pipeline that constructs preference data without human annotation or task-specific reward models. By combining evaluators of content fidelity, textual alignment, and perceptual quality, MPO identifies reliable preference pairs, which are then used to optimize the model via direct preference optimization (DPO). Extensive experiments on both garment generation and virtual dressing benchmarks demonstrate that VersaVogue consistently outperforms existing methods in visual fidelity, semantic consistency, and fine-grained controllability.

cs.CV↗
arXiv2026

Spatio-temporal Decoupled Knowledge Compensator for Few-Shot Action Recognition

Few-Shot Action Recognition (FSAR) is a challenging task that requires recognizing novel action categories with a few labeled videos. Recent works typically apply semantically coarse category names as auxiliary contexts to guide the learning of discriminative visual features. However, such context provided by the action names is too limited to provide sufficient background knowledge for capturing novel spatial and temporal concepts in actions. In this paper, we propose DiST, an innovative Decomposition-incorporation framework for FSAR that makes use of decoupled Spatial and Temporal knowledge provided by large language models to learn expressive multi-granularity prototypes. In the decomposition stage, we decouple vanilla action names into diverse spatio-temporal attribute descriptions (action-related knowledge). Such commonsense knowledge complements semantic contexts from spatial and temporal perspectives. In the incorporation stage, we propose Spatial/Temporal Knowledge Compensators (SKC/TKC) to discover discriminative object-level and frame-level prototypes, respectively. In SKC, object-level prototypes adaptively aggregate important patch tokens under the guidance of spatial knowledge. Moreover, in TKC, frame-level prototypes utilize temporal attributes to assist in inter-frame temporal relation modeling. These learned prototypes thus provide transparency in capturing fine-grained spatial details and diverse temporal patterns. Experimental results show DiST achieves state-of-the-art results on five standard FSAR datasets.

cs.CV↗
arXiv2026

Zero-shot HOI Detection with MLLM-based Detector-agnostic Interaction Recognition

Zero-shot Human-object interaction (HOI) detection aims to locate humans and objects in images and recognize their interactions. While advances in open-vocabulary object detection provide promising solutions for object localization, interaction recognition (IR) remains challenging due to the combinatorial diversity of interactions. Existing methods, including two-stage methods, tightly couple IR with a specific detector and rely on coarse-grained vision-language model (VLM) features, which limit generalization to unseen interactions. In this work, we propose a decoupled framework that separates object detection from IR and leverages multi-modal large language models (MLLMs) for zero-shot IR. We introduce a deterministic generation method that formulates IR as a visual question answering task and enforces deterministic outputs, enabling training-free zero-shot IR. To further enhance performance and efficiency by fine-tuning the model, we design a spatial-aware pooling module that integrates appearance and pairwise spatial cues, and a one-pass deterministic matching method that predicts all candidate interactions in a single forward pass. Extensive experiments on HICO-DET and V-COCO demonstrate that our method achieves superior zero-shot performance, strong cross-dataset generalization, and the flexibility to integrate with any object detectors without retraining. The codes are publicly available at https://github.com/SY-Xuan/DA-HOI.

cs.CV↗
arXiv2026

IMAGAgent: Orchestrating Multi-Turn Image Editing via Constraint-Aware Planning and Reflection

Existing multi-turn image editing paradigms are often confined to isolated single-step execution. Due to a lack of context-awareness and closed-loop feedback mechanisms, they are prone to error accumulation and semantic drift during multi-turn interactions, ultimately resulting in severe structural distortion of the generated images. For that, we propose \textbf{IMAGAgent}, a multi-turn image editing agent framework based on a "plan-execute-reflect" closed-loop mechanism that achieves deep synergy among instruction parsing, tool scheduling, and adaptive correction within a unified pipeline. Specifically, we first present a constraint-aware planning module that leverages a vision-language model (VLM) to precisely decompose complex natural language instructions into a series of executable sub-tasks, governed by target singularity, semantic atomicity, and visual perceptibility. Then, the tool-chain orchestration module dynamically constructs execution paths based on the current image, the current sub-task, and the historical context, enabling adaptive scheduling and collaborative operation among heterogeneous operation models covering image retrieval, segmentation, detection, and editing. Finally, we devise a multi-expert collaborative reflection mechanism where a central large language model (LLM) receives the image to be edited and synthesizes VLM critiques into holistic feedback, simultaneously triggering fine-grained self-correction and recording feedback outcomes to optimize future decisions. Extensive experiments on our constructed \textbf{MTEditBench} and the MagicBrush dataset demonstrate that IMAGAgent achieves performance significantly superior to existing methods in terms of instruction consistency, editing precision, and overall quality. The code is available at https://github.com/hackermmzz/IMAGAgent.git.

cs.GR↗
arXiv2026

DMAP: Human-Aligned Structural Document Map for Multimodal Document Understanding

Existing multimodal document question-answering (QA) systems predominantly rely on flat semantic retrieval, representing documents as a set of disconnected text chunks and largely neglecting their intrinsic hierarchical and relational structures. Such flattening disrupts logical and spatial dependencies - such as section organization, figure-text correspondence, and cross-reference relations, that humans naturally exploit for comprehension. To address this limitation, we introduce a document-level structural Document MAP (DMAP), which explicitly encodes both hierarchical organization and inter-element relationships within multimodal documents. Specifically, we design a Structured-Semantic Understanding Agent to construct DMAP by organizing textual content together with figures, tables, charts, etc. into a human-aligned hierarchical schema that captures both semantic and layout dependencies. Building upon this representation, a Reflective Reasoning Agent performs structure-aware and evidence-driven reasoning, dynamically assessing the sufficiency of retrieved context and iteratively refining answers through targeted interactions with DMAP. Extensive experiments on MMDocQA benchmarks demonstrate that DMAP yields document-specific structural representations aligned with human interpretive patterns, substantially enhancing retrieval precision, reasoning consistency, and multimodal comprehension over conventional RAG-based approaches. Code is available at https://github.com/Forlorin/DMAP

cs.IR↗
arXiv2026

Combating Noisy Labels through Fostering Self- and Neighbor-Consistency

Label noise is pervasive in various real-world scenarios, posing challenges in supervised deep learning. Deep networks are vulnerable to such label-corrupted samples due to the memorization effect. One major stream of previous methods concentrates on identifying clean data for training. However, these methods often neglect imbalances in label noise across different mini-batches and devote insufficient attention to out-of-distribution noisy data. To this end, we propose a noise-robust method named Jo-SNC (\textbf{Jo}int sample selection and model regularization based on \textbf{S}elf- and \textbf{N}eighbor-\textbf{C}onsistency). Specifically, we propose to employ the Jensen-Shannon divergence to measure the ``likelihood'' of a sample being clean or out-of-distribution. This process factors in the nearest neighbors of each sample to reinforce the reliability of clean sample identification. We design a self-adaptive, data-driven thresholding scheme to adjust per-class selection thresholds. While clean samples undergo conventional training, detected in-distribution and out-of-distribution noisy samples are trained following partial label learning and negative learning, respectively. Finally, we advance the model performance further by proposing a triplet consistency regularization that promotes self-prediction consistency, neighbor-prediction consistency, and feature consistency. Extensive experiments on various benchmark datasets and comprehensive ablation studies demonstrate the effectiveness and superiority of our approach over existing state-of-the-art methods.

cs.CV↗
arXiv2025

Spatiotemporal-Untrammelled Mixture of Experts for Multi-Person Motion Prediction

Comprehensively and flexibly capturing the complex spatio-temporal dependencies of human motion is critical for multi-person motion prediction. Existing methods grapple with two primary limitations: i) Inflexible spatiotemporal representation due to reliance on positional encodings for capturing spatiotemporal information. ii) High computational costs stemming from the quadratic time complexity of conventional attention mechanisms. To overcome these limitations, we propose the Spatiotemporal-Untrammelled Mixture of Experts (ST-MoE), which flexibly explores complex spatio-temporal dependencies in human motion and significantly reduces computational cost. To adaptively mine complex spatio-temporal patterns from human motion, our model incorporates four distinct types of spatiotemporal experts, each specializing in capturing different spatial or temporal dependencies. To reduce the potential computational overhead while integrating multiple experts, we introduce bidirectional spatiotemporal Mamba as experts, each sharing bidirectional temporal and spatial Mamba in distinct combinations to achieve model efficiency and parameter economy. Extensive experiments on four multi-person benchmark datasets demonstrate that our approach not only outperforms state-of-art in accuracy but also reduces model parameter by 41.38% and achieves a 3.6x speedup in training. The code is available at https://github.com/alanyz106/ST-MoE.

cs.CV↗
arXiv2025

FoundIR-v2: Optimizing Pre-Training Data Mixtures for Image Restoration Foundation Model

Recent studies have witnessed significant advances in image restoration foundation models driven by improvements in the scale and quality of pre-training data. In this work, we find that the data mixture proportions from different restoration tasks are also a critical factor directly determining the overall performance of all-in-one image restoration models. To this end, we propose a high-capacity diffusion-based image restoration foundation model, FoundIR-v2, which adopts a data equilibrium scheduling paradigm to dynamically optimize the proportions of mixed training datasets from different tasks. By leveraging the data mixing law, our method ensures a balanced dataset composition, enabling the model to achieve consistent generalization and comprehensive performance across diverse tasks. Furthermore, we introduce an effective Mixture-of-Experts (MoE)-driven scheduler into generative pre-training to flexibly allocate task-adaptive diffusion priors for each restoration task, accounting for the distinct degradation forms and levels exhibited by different tasks. Extensive experiments demonstrate that our method can address over 50 sub-tasks across a broader scope of real-world scenarios and achieves favorable performance against state-of-the-art approaches.

cs.CV↗
arXiv2025

PGP-DiffSR: Phase-Guided Progressive Pruning for Efficient Diffusion-based Image Super-Resolution

Although diffusion-based models have achieved impressive results in image super-resolution, they often rely on large-scale backbones such as Stable Diffusion XL (SDXL) and Diffusion Transformers (DiT), which lead to excessive computational and memory costs during training and inference. To address this issue, we develop a lightweight diffusion method, PGP-DiffSR, by removing redundant information from diffusion models under the guidance of the phase information of inputs for efficient image super-resolution. We first identify the intra-block redundancy within the diffusion backbone and propose a progressive pruning approach that removes redundant blocks while reserving restoration capability. We note that the phase information of the restored images produced by the pruned diffusion model is not well estimated. To solve this problem, we propose a phase-exchange adapter module that explores the phase information of the inputs to guide the pruned diffusion model for better restoration performance. We formulate the progressive pruning approach and the phase-exchange adapter module into a unified model. Extensive experiments demonstrate that our method achieves competitive restoration quality while significantly reducing computational load and memory consumption. The code is available at https://github.com/yzb1997/PGP-DiffSR.

cs.CV↗
arXiv2025

fMRI2GES: Co-speech Gesture Reconstruction from fMRI Signal with Dual Brain Decoding Alignment

Understanding how the brain responds to external stimuli and decoding this process has been a significant challenge in neuroscience. While previous studies typically concentrated on brain-to-image and brain-to-language reconstruction, our work strives to reconstruct gestures associated with speech stimuli perceived by brain. Unfortunately, the lack of paired \{brain, speech, gesture\} data hinders the deployment of deep learning models for this purpose. In this paper, we introduce a novel approach, \textbf{fMRI2GES}, that allows training of fMRI-to-gesture reconstruction networks on unpaired data using \textbf{Dual Brain Decoding Alignment}. This method relies on two key components: (i) observed texts that elicit brain responses, and (ii) textual descriptions associated with the gestures. Then, instead of training models in a completely supervised manner to find a mapping relationship among the three modalities, we harness an fMRI-to-text model, a text-to-gesture model with paired data and an fMRI-to-gesture model with unpaired data, establishing dual fMRI-to-gesture reconstruction patterns. Afterward, we explicitly align two outputs and train our model in a self-supervision way. We show that our proposed method can reconstruct expressive gestures directly from fMRI recordings. We also investigate fMRI signals from different ROIs in the cortex and how they affect generation results. Overall, we provide new insights into decoding co-speech gestures, thereby advancing our understanding of neuroscience and cognitive science.

cs.AI↗
arXiv2025

PC-Diffusion: Aligning Diffusion Models with Human Preferences via Preference Classifier

Diffusion models have achieved remarkable success in conditional image generation, yet their outputs often remain misaligned with human preferences. To address this, recent work has applied Direct Preference Optimization (DPO) to diffusion models, yielding significant improvements.~However, DPO-like methods exhibit two key limitations: 1) High computational cost,due to the entire model fine-tuning; 2) Sensitivity to reference model quality}, due to its tendency to introduce instability and bias. To overcome these limitations, we propose a novel framework for human preference alignment in diffusion models (PC-Diffusion), using a lightweight, trainable Preference Classifier that directly models the relative preference between samples. By restricting preference learning to this classifier, PC-Diffusion decouples preference alignment from the generative model, eliminating the need for entire model fine-tuning and reference model reliance.~We further provide theoretical guarantees for PC-Diffusion:1) PC-Diffusion ensures that the preference-guided distributions are consistently propagated across timesteps. 2)The training objective of the preference classifier is equivalent to DPO, but does not require a reference model.3) The proposed preference-guided correction can progressively steer generation toward preference-aligned regions.~Empirical results show that PC-Diffusion achieves comparable preference consistency to DPO while significantly reducing training costs and enabling efficient and stable preference-guided generation.

cs.CV↗
arXiv2025

LeCoT: revisiting network architecture for two-view correspondence pruning

Two-view correspondence pruning aims to accurately remove incorrect correspondences (outliers) from initial ones and is widely applied to various computer vision tasks. Current popular strategies adopt multilayer perceptron (MLP) as the backbone, supplemented by additional modules to enhance the network ability to handle context information, which is a known limitation of MLPs. In contrast, we introduce a novel perspective for capturing correspondence context information without extra design modules. To this end, we design a two-view correspondence pruning network called LeCoT, which can naturally leverage global context information at different stages. Specifically, the core design of LeCoT is the Spatial-Channel Fusion Transformer block, a newly proposed component that efficiently utilizes both spatial and channel global context information among sparse correspondences. In addition, we integrate the proposed prediction block that utilizes correspondence features from intermediate stages to generate a probability set, which acts as guiding information for subsequent learning phases, allowing the network to more effectively capture robust global context information. Notably, this prediction block progressively refines the probability set, thereby mitigating the issue of information loss that is common in the traditional one. Extensive experiments prove that the proposed LeCoT outperforms state-of-the-art methods in correspondence pruning, relative pose estimation, homography estimation, visual localization, and $3$D~reconstruction tasks. The code is provided in https://github.com/Dailuanyuan2024/LeCoT-Revisiting-Network-Architecture-for-Two-View-Correspondence-Pruning.

cs.CV↗
arXiv2025

See the Text: From Tokenization to Visual Reading

People see text. Humans read by recognizing words as visual objects, including their shapes, layouts, and patterns, before connecting them to meaning, which enables us to handle typos, distorted fonts, and various scripts effectively. Modern large language models (LLMs), however, rely on subword tokenization, fragmenting text into pieces from a fixed vocabulary. While effective for high-resource languages, this approach over-segments low-resource languages, yielding long, linguistically meaningless sequences and inflating computation. In this work, we challenge this entrenched paradigm and move toward a vision-centric alternative. Our method, SeeTok, renders text as images (visual-text) and leverages pretrained multimodal LLMs to interpret them, reusing strong OCR and text-vision alignment abilities learned from large-scale multimodal training. Across three different language tasks, SeeTok matches or surpasses subword tokenizers while requiring 4.43 times fewer tokens and reducing FLOPs by 70.5%, with additional gains in cross-lingual generalization, robustness to typographic noise, and linguistic hierarchy. SeeTok signals a shift from symbolic tokenization to human-like visual reading, and takes a step toward more natural and cognitively inspired language models.

cs.CV↗