Embedding-as-a-Service (EaaS) is an effective and convenient deployment solution for addressing various NLP tasks. Nevertheless, recent research has shown that EaaS is vulnerable to model extraction attacks, which could lead to significant economic losses for model providers. For copyright protection, existing methods inject watermark embeddings into text embeddings and use them to detect copyright infringement. However, current watermarking methods often resist only a subset of attacks and fail to provide comprehensive protection. To this end, we present the region-triggered semantic watermarking framework called RegionMarker, which defines trigger regions within a low-dimensional space and injects watermarks into text embeddings associated with these regions. By utilizing a secret dimensionality reduction matrix to project onto this subspace and randomly selecting trigger regions, RegionMarker makes it difficult for watermark removal attacks to evade detection. Furthermore, by embedding watermarks across the entire trigger region and using the text embedding as the watermark, RegionMarker is resilient to both paraphrasing and dimension-perturbation attacks. Extensive experiments on various datasets show that RegionMarker is effective in resisting different attack methods, thereby protecting the copyright of EaaS.

Diffusion models have advanced fine-grained garment generation, yet balancing controllability, efficiency, and texture fidelity remains challenging. Adapter-based methods often yield incoherent details, while full fine-tuning is computationally expensive and prone to overwriting pretrained priors. To address these limitations, we propose IMAGGarment+, an efficient diffusion framework for controllable and high-quality garment synthesis. It comprises two key modules designed for efficient and attribute-aware conditioning. First, we introduce an attribute-wise feature extractor (AFE) that disentangles key garment attributes, silhouette, logo, position, and color, into parallel latent streams. Each stream is optimized independently via LoRA, ensuring minimal parameter overhead while retaining expressive capacity. Second, we develop an attribute-adaptive attention (AA) module to inject attribute-specific cues into the generative process through a selective, layer-wise injection strategy. Specifically, silhouette and color features are injected into early decoder layers to guide structural and appearance formation, while logo features are propagated across all layers to ensure cross-scale consistency. Extensive experiments on fine-grained garment benchmarks demonstrate that IMAGGarment+ outperforms state-of-the-art baselines with less than 20% additional parameters, validating its effectiveness and efficiency.

Zifeng Cheng, Lingyun Qian, Zhiwei Jiang, Cong Wang, Yafeng Yin, Fei Shen, Ao Zhou, Qing Gu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Shufan Yang, Zifeng Cheng, Zhiwei Jiang, Qingfeng Qi, Yafeng Yin, Cong Wang, Ao Zhou, Qing Gu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

This paper presents IMAGGarment, a fine-grained garment generation (FGG) framework that enables high-fidelity garment synthesis with precise control over silhouette, color, and logo placement. Unlike existing methods that are limited to single-condition inputs, IMAGGarment addresses the challenges of multi-conditional controllability in personalized fashion design and digital apparel applications. Specifically, IMAGGarment employs a two-stage training strategy to separately model global appearance and local details, while enabling unified and controllable generation through end-to-end inference. In the first stage, we propose a global appearance model that jointly encodes silhouette and color using a mixed attention module and a color adapter. In the second stage, we present a local enhancement model with an adaptive appearance-aware module to inject user-defined logos and spatial constraints, enabling accurate placement and visual consistency. To support this task, we release GarmentBench, a large-scale dataset comprising over 180 K garment samples paired with multi-level design conditions, including sketches, color references, logo placements, and textual prompts. Extensive experiments demonstrate that our method outperforms existing baselines, achieving superior structural stability, color fidelity, and local controllability performance.

Existing virtual try-on (VTON) methods provide only limited user control over garment attributes and generally overlook essential factors such as face, pose, and scene context. To address these limitations, we introduce the virtual dressing (VD) task, which aims to synthesize freely editable human images conditioned on fixed garments and optional user-defined inputs. We further propose a comprehensive affinity metric index (CAMI) to quantify the consistency between generated outputs and reference garments. We present IMAGDressing-v1, which leverages a garment-specific U-Net to integrate semantic features from CLIP and texture features from a VAE. To incorporate these garment features into a frozen denoising U-Net for flexible text-driven scene control, we employ a hybrid attention mechanism composed of frozen self-attention and trainable cross-attention layers. IMAGDressing-v1 seamlessly integrates with extension modules, such as ControlNet and IP-Adapter, enabling enhanced diversity and controllability. To alleviate data constraints, we introduce the Interactive Garment Pairing (IGPair) dataset, comprising over 300,000 garment–image pairs and a standardized data assembly pipeline. Extensive experiments demonstrate that IMAGDressing-v1 achieves state-of-the-art performance in controlled human image synthesis. The code and model will be available at https://github.com/muzishen/IMAGDressing.

Recent research showcases the considerable potential of conditional diffusion models for generating consistent stories. However, current methods, which primarily generate stories in a caption-dependent manner, often overlook the importance of contextual consistency and the relevance of frames during sequential generation. To address this, we propose a novel Rich-contextual Conditional Diffusion Models (RCDMs), a two-stage approach designed to enhance story generation's semantic consistency and temporal consistency. Specifically, in the first stage, the frame-prior transformer diffusion model is presented to predict the frame semantic embedding of the unknown clip by aligning the semantic correlations between the captions and frames of the known clip. The second stage establishes a robust model with rich contextual conditions, including reference images of the known clip, the predicted frame semantic embedding of the unknown clip, and text embeddings of all captions. By jointly injecting these rich contextual conditions at the image and feature levels, RCDMs can generate semantic and temporal consistency stories. Moreover, RCDMs can generate consistent stories with a single forward inference compared to autoregressive models. Our qualitative and quantitative results demonstrate that our proposed RCDMs outperform in challenging scenarios.

Zifeng Cheng, Zhonghui Wang, Yuchen Fu, Zhiwei Jiang, Yafeng Yin, Cong Wang, Qing Gu. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

The success of the text-guided diffusion model has inspired the development and release of numerous powerful diffusion models within the open-source community. These models are typically fine-tuned on various expert datasets, showcasing diverse denoising capabilities. Leveraging multiple high-quality models to produce stronger generation ability is valuable, but has not been extensively studied. Existing methods primarily adopt parameter merging strategies to produce a new static model. However, they overlook the fact that the divergent denoising capabilities of the models may dynamically change across different states, such as when experiencing different prompts, initial noises, denoising steps, and spatial locations. In this paper, we propose a novel ensembling method, Adaptive Feature Aggregation (AFA), which dynamically adjusts the contributions of multiple models at the feature level according to various states (i.e., prompts, initial noises, denoising steps, and spatial locations), thereby keeping the advantages of multiple diffusion models, while suppressing their disadvantages. Specifically, we design a lightweight Spatial-Aware Block-Wise (SABW) feature aggregator that adaptive aggregates the block-wise intermediate features from multiple U-Net denoisers into a unified one. The core idea lies in dynamically producing an individual attention map for each model's features by comprehensively considering various states. It is worth noting that only SABW is trainable with about 50 million parameters, while other models are frozen. Both the quantitative and qualitative experiments demonstrate the effectiveness of our proposed method.

Large language models (LLMs) have achieved remarkable performance across many generation tasks. Nevertheless, effectively aligning them with desired behaviors remains a significant challenge. Activation steering is an effective and cost-efficient approach that directly modifies the activations of LLMs during the inference stage, aligning their responses with the desired behaviors and avoiding the high cost of fine-tuning. Existing methods typically indiscriminately intervene to all generations or rely solely on the question to determine intervention, which limits the accurate assessment of the intervention strength. To this end, we propose the **F**lexible **A**ctivation **S**teering with **B**acktracking (**FASB**) framework, which dynamically determines both the necessity and strength of intervention by tracking the internal states of the LLMs during generation, considering both the question and the generated content. Since intervening after detecting a deviation from the desired behavior is often too late, we further propose the backtracking mechanism to correct the deviated tokens and steer the LLMs toward the desired behavior. Extensive experiments on the TruthfulQA dataset and six multiple-choice datasets demonstrate that our method outperforms baselines. Our code will be released at https://github.com/gjw185/FASB.

Sign Language Video Generation (SLVG) seeks to generate identity-preserving sign language videos from spoken language texts. Existing methods primarily rely on the single coarse condition (e.g., skeleton sequences) as the intermediary to bridge the translation model and the video generation model, which limits both the naturalness and expressiveness of the generated videos. To overcome these limitations, we propose SignViP, a novel SLVG framework that incorporate multiple fine-grained conditions for improved generation fidelity. Rather than directly translating error-prone high-dimensional conditions, SignViP adopts a discrete tokenization paradigm to integrate and represent fine-grained conditions (i.e., fine-grained poses and 3D hands). SignViP contains three core components. (1) Sign Video Diffusion Model is jointly trained with a multi-condition encoder to learn continuous embeddings that encapsulate fine-grained motion and appearance. (2) Finite Scalar Quantization (FSQ) Autoencoder is further trained to compress and quantize these embeddings into discrete tokens for compact representation of the conditions. (3) Multi-Condition Token Translator is trained to translate spoken language text to discrete multi-condition tokens. During inference, Multi-Condition Token Translator first translates the spoken language text into discrete multi-condition tokens. These tokens are then decoded to continuous embeddings by FSQ Autoencoder, which are subsequently injected into Sign Video Diffusion Model to guide video generation. Experimental results show that SignViP achieves state-of-the-art performance across metrics, including video quality, temporal coherence, and semantic fidelity. The code is available at https://github.com/umnooob/signvip/.

Zero-Shot Cross-Modal Retrieval (ZS-CMR) aims to perform cross-modal retrieval on data of unseen classes, where a key challenge is how to address the modality-gap and domain-shift problems simultaneously. Existing methods tackle this challenge mainly by embracing a sample-label alignment paradigm, which aligns samples of different modalities but of the same class with the word embedding of their class label. However, these methods only focus on the class-level alignment and overlook the alignment of rich fine-grained semantic information in samples, incurring coarse understanding of sample matching and poor generalization on unseen classes. In this article, we propose a novel Fine-Grained Alignment Network, an end-to-end framework that learns representation with two fine-grained alignment strategies, yielding representation space that can be better generalized to unseen classes. Specifically, we extract two kinds of fine-grained representations, region embedding and label distribution, respectively, from aspects of both feature and label. To optimize the region embedding, we propose a Fine-Grained Contrastive Learning (FGCL) strategy to simultaneously conduct class-level alignment and model the intra-class discrepancy. To optimize the label distribution, we propose a Fine-Grained Label Alignment (FGLA) strategy to align diverse fine-grained semantic information among samples, rather than merely label information. Finally, both region embedding and label distribution are utilized together to perform ZS-CMR at a finer granularity. Experimental results on three widely used datasets demonstrate that our method outperforms the state-of-the-art methods by a large margin. Detailed ablation studies have also been carried out, which provably affirm the advantage of each component we propose. Our code will be available at https://github.com/ShipingGe/FGAN .

Recent work has showcased the significant potential of diffusion models in pose-guided person image synthesis. However, owing to the inconsistency in pose between the source and target images, synthesizing an image with a distinct pose, relying exclusively on the source image and target pose information, remains a formidable challenge. This paper presents Progressive Conditional Diffusion Models (PCDMs) that incrementally bridge the gap between person images under the target and source poses through three stages. Specifically, in the first stage, we design a simple prior conditional diffusion model that predicts the global features of the target image by mining the global alignment relationship between pose coordinates and image appearance. Then, the second stage establishes a dense correspondence between the source and target images using the global features from the previous stage, and an inpainting conditional diffusion model is proposed to further align and enhance the contextual features, generating a coarse-grained person image. In the third stage, we propose a refining conditional diffusion model to utilize the coarsely generated image from the previous stage as a condition, achieving texture restoration and enhancing fine-detail consistency. The three-stage PCDMs work progressively to generate the final high-quality and high-fidelity synthesized image. Both qualitative and quantitative results demonstrate the consistency and photorealism of our proposed PCDMs under challenging scenarios. The code and model will be available at https://github.com/tencent-ailab/PCDMs.

Recent advancements in Automatic Prompt Optimization (APO) for text-to-image generation have streamlined user input while ensuring high-quality image output. However, most APO methods are trained assuming a fixed text-to-image model, which is impractical given the emergence of new models. To address this, we propose a novel task, model-generalized automatic prompt optimization (MGAPO), which trains APO methods on a set of known models to enable generalization to unseen models during testing. MGAPO presents significant challenges. First, we experimentally confirm the suboptimal performance of existing APO methods on unseen models. We then introduce a two-stage prompt optimization method, AP-Adapter. In the first stage, a large language model is used to rewrite the prompts. In the second stage, we propose a novel method to construct an enhanced representation space by leveraging inter-model differences. This space captures the characteristics of multiple domain models, storing them as domain prototypes. These prototypes serve as anchors to adjust prompt representations, enabling generalization to unseen models. The optimized prompt representations are subsequently used to generate conditional representations for controllable image generation. We curate a multi-modal, multi-model dataset that includes multiple diffusion models and their corresponding text-image data, and conduct experiments under a model generalization setting. The experimental results demonstrate the AP-Adapter's ability to enable the automatic prompts to generalize well to previously unseen diffusion models, generating high-quality images.

For deep ordinal classification, learning a well-structured feature space specific to ordinal classification is helpful to properly capture the ordinal nature among classes. Intuitively, when Euclidean distance metric is used, an ideal ordinal layout in feature space would be that the sample clusters are arranged in class order along a straight line in space. However, enforcing samples to conform to a specific layout in the feature space is a challenging problem. To address this problem, in this paper, we propose a novel Constrained Proxies Learning (CPL) method, which can learn a proxy for each ordinal class and then adjusts the global layout of classes by constraining these proxies. Specifically, we propose two kinds of strategies: hard layout constraint and soft layout constraint. The hard layout constraint is realized by directly controlling the generation of proxies to force them to be placed in a strict linear layout or semicircular layout (i.e., two instantiations of strict ordinal layout). The soft layout constraint is realized by constraining that the proxy layout should always produce unimodal proxy-to-proxies similarity distribution for each proxy (i.e., to be a relaxed ordinal layout). Experiments show that the proposed CPL method outperforms previous deep ordinal classification methods under the same setting of feature extractor.

Automated Essay Scoring (AES) aims to evaluate the quality score for input essays.In this work, we propose a novel unsupervised AES approach ULRA, which does not require groundtruth scores of essays for training.The core idea of our ULRA is to use multiple heuristic quality signals as the pseudo-groundtruth, and then train a neural AES model by learning from the aggregation of these quality signals.To aggregate these inconsistent quality signals into a unified supervision, we view the AES task as a ranking problem, and design a special Deep Pairwise Rank Aggregation (DPRA) loss for training.In the DPRA loss, we set a learnable confidence weight for each signal to address the conflicts among signals, and train the neural AES model in a pairwise way to disentangle the cascade effect among partialorder pairs.Experiments on eight prompts of ASPA dataset show that ULRA achieves the state-of-the-art performance compared with previous unsupervised methods in terms of both transductive and inductive settings.Further, our approach achieves comparable performance with many existing domain-adapted supervised models, showing the effectiveness of ULRA.The code is available at https: //github.com/tenvence/ulra.

Audio-Visual Event Localization (AVEL) aims to locate events that are both visible and audible in a video. Existing AVEL methods primarily focus on learning generic localization patterns that are applicable to all events. However, events often exhibit modality biases, such as visual-dominated, audio-dominated, or modality-balanced, which can lead to different localization preferences. These preferences may be overlooked by existing methods, resulting in unsatisfactory localization performance. To address this issue, this paper proposes a novel event-aware localization paradigm, which first identifies the event category and then leverages localization preferences specific to that event for improved event localization. To achieve this, we introduce a memory-assisted metric learning framework, which utilizes historic segments as anchors to adjust the unified representation space for both event classification and event localization. To provide sufficient information for this metric learning, we design a spatial-temporal audio-visual fusion encoder to capture the spatial and temporal interaction between audio and visual modalities. Extensive experiments on the public AVE dataset in both fully-supervised and weakly-supervised settings demonstrate the effectiveness of our approach. Code will be released at https://github.com/ShipingGe/AVEL.

Unsupervised readability assessment aims to evaluate the reading difficulty of text without any manually-labeled data for model training. This is a challenging task because the absence of labeled data makes it difficult for the model to understand what readability is. In this paper, we propose a novel framework to Learn a neural model from Weak Readability Signals (LWRS). Instead of relying on labeled data, LWRS utilizes a set of heuristic signals that specialize in describing text readability from different aspects to guide the model in outputting readability scores for ranking. Specifically, to effectively use multiple heuristic weak signals for model training, we build a multi-signal learning model that ranks the unlabeled texts from multiple readability-related aspects based on intra- and inter-signal learning. We also adopt the pairwise ranking paradigm to reduce the cascade coupling among partial-order pairs. Furthermore, we propose identifying the most representative signal based on the batch-level consensus distribution of all signals. This strategy helps identify the predicted signal that is most correlated with readability in the absence of ground-truth labels. We conduct experiments on three public readability assessment datasets. The experimental results demonstrate that our LWRS outperforms each heuristic signal and their combinations significantly, and can even perform comparably with some supervised methods. Additionally, our LWRS trained on one dataset can be effectively transferred to other datasets, including those in other languages, which indicates its good generalization and potential for wide application.

Recognizing emotions from multi-modal data is an emotion recognition task that requires strong multi-modal representation ability. The general approach to this task is to naturally train the representation model on training data without intervention. However, such natural training scheme is prone to modality bias of representation (i.e., tending to over-encode some informative modalities while neglecting other modalities) and data bias of training (i.e., tending to overfit training data). These biases may lead to instability (e.g., performing poorly when the neglected modality is dominant for recognition) and weak generalization (e.g., performing poorly when unseen data is inconsistent with overfitted data) of the model on unseen data. To address these problems, this paper presents two adversarial training strategies to learn more robust multi-modal representation for multi-label emotion recognition. Firstly, we propose an adversarial temporal masking strategy, which can enhance the encoding of other modalities by masking the most emotion-related temporal units (e.g., words for text or frames for video) of the informative modality. Secondly, we propose an adversarial parameter perturbation strategy, which can enhance the generalization of the model by adding the adversarial perturbation to the parameters of model. Both strategies boost model performance on the benchmark MMER datasets CMU-MOSEI and NEMu. Experimental results demonstrate the effectiveness of the proposed method compared with the previous state-of-the-art method. Code will be released at https://github.com/ShipingGe/MMER.

Emotion-cause pair extraction (ECPE) is a recently proposed task that aims to extract the potential clause pairs of emotions and its corresponding causes in a document. In this article, we propose a new paradigm for the ECPE task. We cast the task as a two-turn machine reading comprehension (MRC) task, i.e., the extraction of emotions and causes is transformed to the task of identifying answer clauses from the input document specific to a query. This two-turn MRC formalization brings several key advantages: First, the QA manner provides an explicit pairing way to identify causes specific to the target emotion; second, it provides a natural way of jointly modeling the emotion extraction, the cause extraction, and the pairing of emotion and cause; and third, it allows us to exploit the well-developed MRC models. Based on the two-turn MRC formalization, we propose a dual-MRC framework to extract emotion-cause pairs in a dual-direction way, which enables a more comprehensive coverage of all pairing cases. Furthermore, we propose a consistent training strategy for the second-turn query, so the model is able to filter the errors produced by the first turn at inference. Experiments on two benchmark datasets demonstrate that our method outperforms previous methods and achieves state-of-the-art performance. All the code and data of this work can be obtained at https://github.com/zifengcheng/CD-MRC .

Open intent classification is a practical yet challenging task in dialogue systems. Its objective is to accurately classify samples of known intents while at the same time detecting those of open (unknown) intents. Existing methods usually use outlier detection algorithms combined with K-class classifier to detect open intents, where K represents the class number of known intents. Different from them, in this paper, we consider another way without using outlier detection algorithms. Specifically, we directly train a (K+1)-class classifier for open intent classification, where the (K+1)-th class represents open intents. To address the challenge that training a (K+1)-class classifier with training samples of only K classes, we propose a deep model based on Soft Labeling and Manifold Mixup (SLMM). In our method, soft labeling is used to reshape the label distribution of the known intent samples, aiming at reducing model’s overconfident on known intents. Manifold mixup is used to generate pseudo samples for open intents, aiming at well optimizing the decision boundary of open intents. Experiments on four benchmark datasets demonstrate that our method outperforms previous methods and achieves state-of-the-art performance. All the code and data of this work can be obtained at.1