Xing Wu

Yeyong Yu, Wenya Hu, Xing Wu et al.

As candidate generation and high-throughput experimentation advance, the primary bottleneck in materials discovery is shifting from property prediction to making reliable evaluations among massive candidate sets. We propose a Knowledge-Augmented Preference Signals Framework, MaterEval, that automatically produces, for the same candidate, two evaluations: an informed judgment that follows expert rules and provides supporting evidence, and a rule-removed blind guess. By pairing the two evaluations as preference data, we guide general-purpose large language models (LLMs), originally lacking materials-specific criteria, from intuitive judgment toward reliable evaluation supported by explicit evidence. To balance throughput, cost, and reliability, we further introduce a fast-slow reasoning scheme that decouples large-scale rapid screening from in-depth review on a small subset. Using high-entropy alloy (HEA) assessment as a case study, we show that, without external retrieval and relying solely on internalized capabilities, small open-source LLMs achieve substantial gains in accuracy, conclusion consistency, and evidence discrimination, approaching the performance of rule-based closed-source LLMs. These results demonstrate that expert rules can be systematically transformed into learnable preference signals, enabling a low-cost and deployable evaluation module for autonomous materials discovery loops.

Xing Wu, Chaochen Gao, Zijia Lin et al.

Contrastive learning has been extensively studied in sentence embedding learning, which assumes that the embeddings of different views of the same sentence are closer. The constraint brought by this assumption is weak, and a good sentence representation should also be able to reconstruct the original sentence fragments. Therefore, this paper proposes an information-aggregated contrastive learning framework for learning unsupervised sentence embeddings, termed InfoCSE. InfoCSE forces the representation of [CLS] positions to aggregate denser sentence information by introducing an additional Masked language model task and a well-designed network. We evaluate the proposed InfoCSE on several benchmark datasets w.r.t the semantic text similarity (STS) task. Experimental results show that InfoCSE outperforms SimCSE by an average Spearman correlation of 2.60% on BERT-base, and 1.77% on BERT-large, achieving state-of-the-art results among unsupervised sentence representation learning methods. Our code are available at https://github.com/caskcsg/sentemb/tree/main/InfoCSE.

Xing Wu, Guangyuan Ma, Meng Lin et al.

Dense passage retrieval aims to retrieve the relevant passages of a query from a large corpus based on dense representations (i.e., vectors) of the query and the passages. Recent studies have explored improving pre-trained language models to boost dense retrieval performance. This paper proposes CoT-MAE (ConTextual Masked Auto-Encoder), a simple yet effective generative pre-training method for dense passage retrieval. CoT-MAE employs an asymmetric encoder-decoder architecture that learns to compress the sentence semantics into a dense vector through self-supervised and context-supervised masked auto-encoding. Precisely, self-supervised masked auto-encoding learns to model the semantics of the tokens inside a text span, and context-supervised masked auto-encoding learns to model the semantical correlation between the text spans. We conduct experiments on large-scale passage retrieval benchmarks and show considerable improvements over strong baselines, demonstrating the high efficiency of CoT-MAE. Our code is available at https://github.com/caskcsg/ir/tree/main/cotmae.

Xing Wu, Chaochen Gao, Zijia Lin et al.

Video language pre-training methods have mainly adopted sparse sampling techniques to alleviate the temporal redundancy of videos. Though effective, sparse sampling still suffers inter-modal redundancy: visual redundancy and textual redundancy. Compared with highly generalized text, sparsely sampled frames usually contain text-independent portions, called visual redundancy. Sparse sampling is also likely to miss important frames corresponding to some text portions, resulting in textual redundancy. Inter-modal redundancy leads to a mismatch of video and text information, hindering the model from better learning the shared semantics across modalities. To alleviate it, we propose Redundancy-aware Video-language Pre-training. We design a redundancy measurement of video patches and text tokens by calculating the cross-modal minimum dis-similarity. Then, we penalize the highredundant video patches and text tokens through a proposed redundancy-aware contrastive learning. We evaluate our method on four benchmark datasets, MSRVTT, MSVD, DiDeMo, and LSMDC, achieving a significant improvement over the previous stateof-the-art results. Our code are available at https://github.com/caskcsg/VLP/tree/main/RaP.

Xing Wu, Guangyuan Ma, Wanhui Qian et al.

Recently, methods have been developed to improve the performance of dense passage retrieval by using context-supervised pre-training. These methods simply consider two passages from the same document to be relevant, without taking into account the possibility of weakly correlated pairs. Thus, this paper proposes query-as-context pre-training, a simple yet effective pre-training technique to alleviate the issue. Query-as-context pre-training assumes that the query derived from a passage is more likely to be relevant to that passage and forms a passage-query pair. These passage-query pairs are then used in contrastive or generative context-supervised pre-training. The pre-trained models are evaluated on large-scale passage retrieval benchmarks and out-of-domain zero-shot benchmarks. Experimental results show that query-as-context pre-training brings considerable gains and meanwhile speeds up training, demonstrating its effectiveness and efficiency. Our code will be available at https://github.com/caskcsg/ir/tree/main/cotmae-qc .

Xing Wu, Qiulian Fang

Cancer survival prediction is important for developing personalized treatments and inducing disease-causing mechanisms. Multi-omics data integration is attracting widespread interest in cancer research for providing information for understanding cancer progression at multiple genetic levels. Many works, however, are limited because of the high dimensionality and heterogeneity of multi-omics data. In this paper, we propose a novel method to integrate multi-omics data for cancer survival prediction, called Stacked AutoEncoder-based Survival Prediction Neural Network (SAEsurv-net). In the cancer survival prediction for TCGA cases, SAEsurv-net addresses the curse of dimensionality with a two-stage dimensionality reduction strategy and handles multi-omics heterogeneity with a stacked autoencoder model. The two-stage dimensionality reduction strategy achieves a balance between computation complexity and information exploiting. The stacked autoencoder model removes most heterogeneities such as data's type and size in the first group of autoencoders, and integrates multiple omics data in the second autoencoder. The experiments show that SAEsurv-net outperforms models based on a single type of data as well as other state-of-the-art methods.

Zhenpeng Su, Xing Wu, Wei Zhou et al.

Dialogue response selection aims to select an appropriate response from several candidates based on a given user and system utterance history. Most existing works primarily focus on post-training and fine-tuning tailored for cross-encoders. However, there are no post-training methods tailored for dense encoders in dialogue response selection. We argue that when the current language model, based on dense dialogue systems (such as BERT), is employed as a dense encoder, it separately encodes dialogue context and response, leading to a struggle to achieve the alignment of both representations. Thus, we propose Dial-MAE (Dialogue Contextual Masking Auto-Encoder), a straightforward yet effective post-training technique tailored for dense encoders in dialogue response selection. Dial-MAE uses an asymmetric encoder-decoder architecture to compress the dialogue semantics into dense vectors, which achieves better alignment between the features of the dialogue context and response. Our experiments have demonstrated that Dial-MAE is highly effective, achieving state-of-the-art performance on two commonly evaluated benchmarks.

Guangyuan Ma, Hongtao Liu, Xing Wu et al.

News recommendation aims to predict click behaviors based on user behaviors. How to effectively model the user representations is the key to recommending preferred news. Existing works are mostly focused on improvements in the supervised fine-tuning stage. However, there is still a lack of PLM-based unsupervised pre-training methods optimized for user representations. In this work, we propose an unsupervised pre-training paradigm with two tasks, i.e. user behavior masking and user behavior generation, both towards effective user behavior modeling. Firstly, we introduce the user behavior masking pre-training task to recover the masked user behaviors based on their contextual behaviors. In this way, the model could capture a much stronger and more comprehensive user news reading pattern. Besides, we incorporate a novel auxiliary user behavior generation pre-training task to enhance the user representation vector derived from the user encoder. We use the above pre-trained user modeling encoder to obtain news and user representations in downstream fine-tuning. Evaluations on the real-world news benchmark show significant performance improvements over existing baselines.

Guangyuan Ma, Xing Wu, Peng Wang et al.

In this paper, we systematically study the potential of pre-training with Large Language Model(LLM)-based document expansion for dense passage retrieval. Concretely, we leverage the capabilities of LLMs for document expansion, i.e. query generation, and effectively transfer expanded knowledge to retrievers using pre-training strategies tailored for passage retrieval. These strategies include contrastive learning and bottlenecked query generation. Furthermore, we incorporate a curriculum learning strategy to reduce the reliance on LLM inferences. Experimental results demonstrate that pre-training with LLM-based document expansion significantly boosts the retrieval performance on large-scale web-search tasks. Our work shows strong zero-shot and out-of-domain retrieval abilities, making it more widely applicable for retrieval when initializing with no human-labeled data.

Guangyuan Ma, Xing Wu, Peng Wang et al.

Passage retrieval aims to retrieve relevant passages from large collections of the open-domain corpus. Contextual Masked Auto-Encoding has been proven effective in representation bottleneck pre-training of a monolithic dual-encoder for passage retrieval. Siamese or fully separated dual-encoders are often adopted as basic retrieval architecture in the pre-training and fine-tuning stages for encoding queries and passages into their latent embedding spaces. However, simply sharing or separating the parameters of the dual-encoder results in an imbalanced discrimination of the embedding spaces. In this work, we propose to pre-train Contextual Masked Auto-Encoder with Mixture-of-Textual-Experts (CoT-MoTE). Specifically, we incorporate textual-specific experts for individually encoding the distinct properties of queries and passages. Meanwhile, a shared self-attention layer is still kept for unified attention modeling. Results on large-scale passage retrieval benchmarks show steady improvement in retrieval performances. The quantitive analysis also shows a more balanced discrimination of the latent embedding spaces.

Xing Wu, Guangyuan Ma, Peng Wang et al.

Growing techniques have been emerging to improve the performance of passage retrieval. As an effective representation bottleneck pretraining technique, the contextual masked auto-encoder utilizes contextual embedding to assist in the reconstruction of passages. However, it only uses a single auto-encoding pre-task for dense representation pre-training. This study brings multi-view modeling to the contextual masked auto-encoder. Firstly, multi-view representation utilizes both dense and sparse vectors as multi-view representations, aiming to capture sentence semantics from different aspects. Moreover, multiview decoding paradigm utilizes both autoencoding and auto-regressive decoders in representation bottleneck pre-training, aiming to provide both reconstructive and generative signals for better contextual representation pretraining. We refer to this multi-view pretraining method as CoT-MAE v2. Through extensive experiments, we show that CoT-MAE v2 is effective and robust on large-scale passage retrieval benchmarks and out-of-domain zero-shot benchmarks.

Zhenpeng Su, Zijia Lin, Xue Bai et al.

Scaling the size of a model enhances its capabilities but significantly increases computation complexity. Mixture-of-Experts models (MoE) address the issue by allowing model size to scale up without substantially increasing training or inference costs. In MoE, there is an important module called the router, which is used to distribute each token to the experts. Currently, the mainstream routing methods include dynamic routing and fixed routing. Despite their promising results, MoE models encounter several challenges. Primarily, for dynamic routing methods, the dispersion of training tokens across multiple experts can lead to underfitting, particularly for infrequent tokens. Additionally, though fixed routing methods can mitigate that issue, they compromise on the diversity of representations. In this paper, we propose \textbf{MaskMoE}, a method designed to enhance token-level learning by employing a routing \textbf{mask}ing technique within the \textbf{M}ixture-\textbf{o}f-\textbf{E}xperts model. MaskMoE is capable of maintaining representation diversity while achieving more comprehensive training. Experimental results demonstrate that our method outperforms previous dominant Mixture-of-Experts models in terms of both perplexity (PPL) and downstream task performance.

Zhenpeng Su, Xing Wu, Xue Bai et al.

Generative language models are usually pretrained on large text corpus via predicting the next token (i.e., sub-word/word/phrase) given the previous ones. Recent works have demonstrated the impressive performance of large generative language models on downstream tasks. However, existing generative language models generally neglect an inherent challenge in text corpus during training, i.e., the imbalance between frequent tokens and infrequent ones. It can lead a language model to be dominated by common and easy-to-learn tokens, thereby overlooking the infrequent and difficult-to-learn ones. To alleviate that, we propose a MiLe Loss function for mitigating the bias of learning difficulties with tokens. During training, it can dynamically assess the learning difficulty of a to-be-learned token, according to the information entropy of the corresponding predicted probability distribution over the vocabulary. Then it scales the training loss adaptively, trying to lead the model to focus more on the difficult-to-learn tokens. On the Pile dataset, we train generative language models at different scales of 468M, 1.2B, and 6.7B parameters. Experiments reveal that models incorporating the proposed MiLe Loss can gain consistent performance improvement on downstream benchmarks.

Zhenpeng Su, Xing Wu, Wei Zhou et al.

ChatGPT has garnered significant interest due to its impressive performance; however, there is growing concern about its potential risks, particularly in the detection of AI-generated content (AIGC), which is often challenging for untrained individuals to identify. Current datasets used for detecting ChatGPT-generated text primarily focus on question-answering tasks, often overlooking tasks with semantic-invariant properties, such as summarization, translation, and paraphrasing. In this paper, we demonstrate that detecting model-generated text in semantic-invariant tasks is more challenging. To address this gap, we introduce a more extensive and comprehensive dataset that incorporates a wider range of tasks than previous work, including those with semantic-invariant properties. In addition, instruction fine-tuning has demonstrated superior performance across various tasks. In this paper, we explore the use of instruction fine-tuning models for detecting text generated by ChatGPT.

Haotian Xu, Xing Wu, Weinong Wang et al.

Can scaling transform reasoning? In this work, we explore the untapped potential of scaling Long Chain-of-Thought (Long-CoT) data to 1000k samples, pioneering the development of a slow-thinking model, RedStar. Through extensive experiments with various LLMs and different sizes, we uncover the ingredients for specialization and scale for Long-CoT training. Surprisingly, even smaller models show significant performance gains with limited data, revealing the sample efficiency of Long-CoT and the critical role of sample difficulty in the learning process. Our findings demonstrate that Long-CoT reasoning can be effectively triggered with just a few thousand examples, while larger models achieve unparalleled improvements. We also introduce reinforcement learning (RL)-scale training as a promising direction for advancing slow-thinking systems. RedStar shines across domains: on the MATH-Hard benchmark, RedStar-code-math boosts performance from 66.2\% to 81.6\%, and on the USA Math Olympiad (AIME), it solves 46.7\% of problems using only 21k mixed-code-math datasets. In multimodal tasks like GeoQA and MathVista-GEO, RedStar-Geo achieves competitive results with minimal Long-CoT data, outperforming other slow-thinking systems like QvQ-Preview. Compared to QwQ, RedStar strikes the perfect balance between reasoning and generalizability. Our work highlights that, with careful tuning, scaling Long-CoT can unlock extraordinary reasoning capabilities-even with limited dataset and set a new standard for slow-thinking models across diverse challenges. Our data and models are released at https://huggingface.co/RedStar-Reasoning.

Yike Guo, Qifeng Liu, Jie Chen et al.

This report presents a comprehensive view of our vision on the development path of the human-machine symbiotic art creation. We propose a classification of the creative system with a hierarchy of 5 classes, showing the pathway of creativity evolving from a mimic-human artist (Turing Artists) to a Machine artist in its own right. We begin with an overview of the limitations of the Turing Artists then focus on the top two-level systems, Machine Artists, emphasizing machine-human communication in art creation. In art creation, it is necessary for machines to understand humans' mental states, including desires, appreciation, and emotions, humans also need to understand machines' creative capabilities and limitations. The rapid development of immersive environment and further evolution into the new concept of metaverse enable symbiotic art creation through unprecedented flexibility of bi-directional communication between artists and art manifestation environments. By examining the latest sensor and XR technologies, we illustrate the novel way for art data collection to constitute the base of a new form of human-machine bidirectional communication and understanding in art creation. Based on such communication and understanding mechanisms, we propose a novel framework for building future Machine artists, which comes with the philosophy that a human-compatible AI system should be based on the "human-in-the-loop" principle rather than the traditional "end-to-end" dogma. By proposing a new form of inverse reinforcement learning model, we outline the platform design of machine artists, demonstrate its functions and showcase some examples of technologies we have developed. We also provide a systematic exposition of the ecosystem for AI-based symbiotic art form and community with an economic model built on NFT technology. Ethical issues for the development of machine artists are also discussed.

Zhiyuan Fan, Weinong Wang, Xing Wu et al.

The evaluation paradigm of LLM-as-judge gains popularity due to its significant reduction in human labor and time costs. This approach utilizes one or more large language models (LLMs) to assess the quality of outputs from other LLMs. However, existing methods rely on generic scoring rubrics that fail to consider the specificities of each question and its problem-solving process, compromising precision and stability in assessments. Inspired by human examination scoring processes, we propose a new evaluation paradigm based on self-adaptive rubrics. Specifically, we create detailed scoring rubrics for each question, capturing the primary and secondary criteria in a structured format of scoring and deduction points that mimic a human evaluator's analytical process. Building on this paradigm, we further develop a novel benchmark called SedarEval, which covers a range of domains including long-tail knowledge, mathematics, coding, and logical reasoning. SedarEval consists of 1,000 meticulously crafted questions, each with its own self-adaptive rubric. To further streamline the evaluation, we train a specialized evaluator language model (evaluator LM) to supplant human graders. Using the same training data, our evaluator LM achieves a higher concordance rate with human grading results than other paradigms, including GPT-4, highlighting the superiority and efficiency of our approach. We release our dataset at https://github.com/wwn1233/sedareval.

Gheorghe Comanici, Eric Bieber, Mike Schaekermann et al. · amazon-science, baidu

In this report, we introduce the Gemini 2.X model family: Gemini 2.5 Pro and Gemini 2.5 Flash, as well as our earlier Gemini 2.0 Flash and Flash-Lite models. Gemini 2.5 Pro is our most capable model yet, achieving SoTA performance on frontier coding and reasoning benchmarks. In addition to its incredible coding and reasoning skills, Gemini 2.5 Pro is a thinking model that excels at multimodal understanding and it is now able to process up to 3 hours of video content. Its unique combination of long context, multimodal and reasoning capabilities can be combined to unlock new agentic workflows. Gemini 2.5 Flash provides excellent reasoning abilities at a fraction of the compute and latency requirements and Gemini 2.0 Flash and Flash-Lite provide high performance at low latency and cost. Taken together, the Gemini 2.X model generation spans the full Pareto frontier of model capability vs cost, allowing users to explore the boundaries of what is possible with complex agentic problem solving.

Pengwei Tang, Yong Liu, Dongjie Zhang et al.

Low-Rank Adaptation (LoRA) is the leading parameter-efficient fine-tuning method for Large Language Models (LLMs). However, the fine-tuned LLMs encounter the issue of catastrophic forgetting of the pre-trained world knowledge. To address this issue, inspired by theoretical insights of null space, we propose LoRA-Null, i.e., Low-Rank Adaptation via null space, which builds adapters initialized from the null space of the pre-trained knowledge activation. Concretely, we randomly collect a few data samples and capture their activations after passing through the LLM layer. We perform Singular Value Decomposition on the input activations to obtain their null space. We use the projection of the pre-trained weights onto the null space as the initialization for adapters. Experimental results demonstrate that this initialization approach can effectively preserve the original pre-trained world knowledge of the LLMs during fine-tuning. Additionally, if we freeze the values of the down-projection matrices during fine-tuning, it achieves even better preservation of the pre-trained world knowledge. LoRA-Null effectively preserves pre-trained world knowledge while maintaining strong fine-tuning performance, as validated by extensive experiments on LLaMA series (LLaMA2, LLaMA3, LLaMA3.1, and LLaMA3.2) across Code, Math, and Instruction Following tasks. We also provide a theoretical guarantee for the capacity of LoRA-Null to retain pre-trained knowledge. Code is in https://github.com/HungerPWAY/LoRA-Null.

Shangpin Peng, Weinong Wang, Zhuotao Tian et al.

Direct Preference Optimization (DPO) has become a cornerstone of reinforcement learning from human feedback (RLHF) due to its simplicity and efficiency. However, existing DPO-based approaches typically treat all preference pairs uniformly, ignoring critical variations in their inherent quality and learning utility, leading to suboptimal data utilization and performance. To address this challenge, we propose Omni-DPO, a dual-perspective optimization framework that jointly accounts for (1) the inherent quality of each preference pair and (2) the model's evolving performance on those pairs. By adaptively weighting samples according to both data quality and the model's learning dynamics during training, Omni-DPO enables more effective training data utilization and achieves better performance. Experimental results on various models and benchmarks demonstrate the superiority and generalization capabilities of Omni-DPO. On textual understanding tasks, Gemma-2-9b-it finetuned with Omni-DPO beats the leading LLM, Claude 3 Opus, by a significant margin of 6.7 points on the Arena-Hard benchmark. On mathematical reasoning tasks, Omni-DPO consistently outperforms the baseline methods across all benchmarks, providing strong empirical evidence for the effectiveness and robustness of our approach. Code and models will be available at https://github.com/pspdada/Omni-DPO.

Shichao Ma, Zhiyuan Ma, Ming Yang et al.

Multi-turn tool-integrated reasoning enables Large Language Models (LLMs) to solve complex tasks through iterative information retrieval. However, current reinforcement learning (RL) frameworks for search-augmented reasoning predominantly rely on sparse outcome-level rewards, leading to a "Double Homogenization Dilemma." This manifests as (1) Process homogenization, where the thinking, reasoning, and tooling involved in generation are ignored. (2) Intra-group homogenization, coarse-grained outcome rewards often lead to inefficiencies in intra-group advantage estimation with methods like Group Relative Policy Optimization (GRPO) during sampling. To address this, we propose Turn-level Stage-aware Policy Optimization (TSPO). TSPO introduces the First-Occurrence Latent Reward (FOLR) mechanism, allocating partial rewards to the step where the ground-truth answer first appears, thereby preserving process-level signals and increasing reward variance within groups without requiring external reward models or any annotations. Extensive experiments demonstrate that TSPO significantly outperforms state-of-the-art baselines, achieving average performance gains of 24% and 13.6% on Qwen2.5-3B and 7B models, respectively.

Yifei Gong, Xing Wu, Wenda Liu et al.

Computer-Aided Design (CAD) is an expert-level task that relies on long-horizon reasoning and coherent modeling actions. Large Language Models (LLMs) have shown remarkable advancements in enabling language agents to tackle real-world tasks. Notably, there has been no investigation into how tool-using LLMs optimally interact with CAD engines, hindering the emergence of LLM-based agentic text-to-CAD modeling systems. We propose ToolCAD, a novel agentic CAD framework deploying LLMs as tool-using agents for text-to-CAD generation. Furthermore, we introduce an interactive CAD modeling gym to rollout reasoning and tool-augmented interaction trajectories with the CAD engine, incorporating hybrid feedback and human supervision. Meanwhile, an end-to-end post-training strategy is presented to enable the LLM agent to elicit refined CAD Modeling Chain of Thought (CAD-CoT) and evolve into proficient CAD tool-using agents via online curriculum reinforcement learning. Our findings demonstrate ToolCAD fills the gap in adopting and training open-source LLMs for CAD tool-using agents, enabling them to perform comparably to proprietary models, paving the way for more accessible and robust autonomous text-to-CAD modeling systems.

Ming Deng, Sijin Sun, Zihao Li et al.

Camouflaged Object Detection (COD) is challenging due to the strong similarity between camouflaged objects and their surroundings, which complicates identification. Existing methods mainly rely on spatial local features, failing to capture global information, while Transformers increase computational costs. To address this, the Frequency-Assisted Mamba-Like Linear Attention Network (FMNet) is proposed, which leverages frequency-domain learning to efficiently capture global features and mitigate ambiguity between objects and the background. FMNet introduces the Multi-Scale Frequency-Assisted Mamba-Like Linear Attention (MFM) module, integrating frequency and spatial features through a multi-scale structure to handle scale variations while reducing computational complexity. Additionally, the Pyramidal Frequency Attention Extraction (PFAE) module and the Frequency Reverse Decoder (FRD) enhance semantics and reconstruct features. Experimental results demonstrate that FMNet outperforms existing methods on multiple COD datasets, showcasing its advantages in both performance and efficiency. Code available at https://github.com/Chranos/FMNet.

Zhiyuan Fan, Tinghao Yu, Yuanjun Cai et al.

Terminal agents have demonstrated strong potential for autonomous command-line execution, yet their training remains constrained by the scarcity of high-quality and diverse execution trajectories. Existing approaches mitigate this bottleneck by synthesizing large-scale terminal task instances for trajectory sampling. However, they primarily focus on scaling the number of tasks while providing limited control over the diversity of execution trajectories that agents actually experience during training. In this paper, we present SkillSynth, an automated framework for terminal task synthesis built on a scenario-mediated skill graph. SkillSynth first constructs a large-scale skill graph, where scenarios serve as intermediate transition nodes that connect diverse command-line skills. It then samples paths from this graph as abstractions of real-world workflows, and uses a multi-agent harness to instantiate them into executable task instances. By grounding task synthesis in graph-sampled workflow paths, SkillSynth explicitly controls the diversity of minimal execution trajectories required to solve the synthesized tasks. Experiments on Terminal-Bench demonstrate the effectiveness of SkillSynth. Moreover, task instances synthesized by SkillSynth have been adopted to train Hy3 Preview, contributing to its enhanced agentic capabilities in terminal-based settings.

Chaochen Gao, Xing Wu, Zijia Lin et al.

Large language models (LLMs) with extended context windows have made significant strides yet remain a challenge due to the scarcity of long documents. Existing methods tend to synthesize long-context data but lack a clear mechanism to reinforce the long-range dependency modeling. To address this limitation, we propose NExtLong, a novel framework for synthesizing long-context data through Negative document Extension. NExtLong decomposes a document into multiple meta-chunks and extends the context by interleaving hard negative distractors retrieved from pretraining corpora. This approach compels the model to discriminate long-range dependent context from distracting content, enhancing its ability to model long-range dependencies. Extensive experiments demonstrate that NExtLong achieves significant performance improvements on the HELMET and RULER benchmarks compared to existing long-context synthesis approaches and leading models, which are trained on non-synthetic long documents. These findings highlight NExtLong's ability to reduce reliance on non-synthetic long documents, making it an effective framework for developing advanced long-context LLMs.

Zhenpeng Su, Xing Wu, Zijia Lin et al.

Large language models (LLM) have been attracting much attention from the community recently, due to their remarkable performance in all kinds of downstream tasks. According to the well-known scaling law, scaling up a dense LLM enhances its capabilities, but also significantly increases the computational complexity. Mixture-of-Experts (MoE) models address that by allowing the model size to grow without substantially raising training or inference costs. Yet MoE models face challenges regarding knowledge sharing among experts, making their performance somehow sensitive to routing accuracy. To tackle that, previous works introduced shared experts and combined their outputs with those of the top $K$ routed experts in an ``addition'' manner. In this paper, inspired by collective matrix factorization to learn shared knowledge among data, we propose CartesianMoE, which implements more effective knowledge sharing among experts in more like a ``multiplication'' manner. Extensive experimental results indicate that CartesianMoE outperforms previous MoE models for building LLMs, in terms of both perplexity and downstream task performance. And we also find that CartesianMoE achieves better expert routing robustness.

Jiang Zhou, Yunhao Wang, Xing Wu et al.

Synthetic data rephrasing has emerged as a powerful technique for enhancing knowledge acquisition during large language model (LLM) pretraining. However, existing approaches operate at the single-document level, rewriting individual web pages in isolation. This confines synthesized examples to intra-document knowledge, missing cross-document relationships and leaving facts with limited associative context. We propose WRAP++ (Web discoveRy Amplified Pretraining), which amplifies the associative context of factual knowledge by discovering cross-document relationships from web hyperlinks and synthesizing joint QA over each discovered document pair. Concretely, WRAP++ discovers high-confidence relational motifs including dual-links and co-mentions, and synthesizes QA that requires reasoning across both documents. This produces relational knowledge absent from either source document alone, creating diverse entry points to the same facts. Because the number of valid entity pairs grows combinatorially, this discovery-driven synthesis also amplifies data scale far beyond single-document rewriting. Instantiating WRAP++ on Wikipedia, we amplify ~8.4B tokens of raw text into 80B tokens of cross-document QA data. On SimpleQA, OLMo-based models at both 7B and 32B scales trained with WRAP++ substantially outperform single-document approaches and exhibit sustained scaling gains, underscoring the advantage of cross-document knowledge discovery and amplification.

Hengzhu Tang, Zefeng Zhang, Zhiping Li et al.

Video Quality Assessment (VQA) is vital for large-scale video retrieval systems, aimed at identifying quality issues to prioritize high-quality videos. In industrial systems, low-quality video characteristics fall into four categories: visual-related issues like mosaics and black boxes, textual issues from video titles and OCR content, and semantic issues like frame incoherence and frame-text mismatch from AI-generated videos. Despite their prevalence in industrial settings, these low-quality videos have been largely overlooked in academic research, posing a challenge for accurate identification. To address this, we introduce the Multi-Branch Collaborative Network (MBCN) tailored for industrial video retrieval systems. MBCN features four branches, each designed to tackle one of the aforementioned quality issues. After each branch independently scores videos, we aggregate these scores using a weighted approach and a squeeze-and-excitation mechanism to dynamically address quality issues across different scenarios. We implement point-wise and pair-wise optimization objectives to ensure score stability and reasonableness. Extensive offline and online experiments on a world-level video search engine demonstrate MBCN's effectiveness in identifying video quality issues, significantly enhancing the retrieval system's ranking performance. Detailed experimental analyses confirm the positive contribution of all four evaluation branches. Furthermore, MBCN significantly improves recognition accuracy for low-quality AI-generated videos compared to the baseline.

Junlong Jia, Ziyang Chen, Xing Wu et al.

Extending LLM context windows is hindered by scarce high-quality long-context data. Recent methods synthesize data with genuine long-range dependencies via information-theoretic verification, selecting contexts that reduce a base model's predictive entropy. However, their single-pass offline construction with a fixed model creates a fundamental off-policy gap: the static screening landscape misaligns with the model's evolving capabilities, causing the training distribution to drift. We propose PolicyLong, shifting data construction towards a dynamic on-policy paradigm. By iteratively re-executing data screening (entropy computation, retrieval, and verification) using the current model, PolicyLong ensures the training distribution tracks evolving capabilities, yielding an emergent self-curriculum. Crucially, both positive and hard negative contexts derive from the current model's entropy landscape, co-evolving what the model learns to exploit and resist. Experiments on RULER, HELMET, and LongBench-v2 (Qwen2.5-3B) show PolicyLong consistently outperforms EntropyLong and NExtLong, with gains growing at longer contexts (e.g., +2.54 at 128K on RULER), confirming the value of on-policy data evolution.

Chaochen Gao, Xing Wu, Zijia Lin et al.

High-quality long-context instruction data is essential for aligning long-context large language models (LLMs). Despite the public release of models like Qwen and Llama, their long-context instruction data remains proprietary. Human annotation is costly and challenging, while template-based synthesis methods limit scale, diversity, and quality. We introduce LongMagpie, a self-synthesis framework that automatically generates large-scale long-context instruction data. Our key insight is that aligned long-context LLMs, when presented with a document followed by special tokens preceding a user turn, auto-regressively generate contextually relevant queries. By harvesting these document-query pairs and the model's responses, LongMagpie produces high-quality instructions without human effort. Experiments on HELMET, RULER, and Longbench v2 demonstrate that LongMagpie achieves leading performance on long-context tasks while maintaining competitive performance on short-context tasks, establishing it as a simple and effective approach for open, diverse, and scalable long-context instruction data synthesis.

Junlong Jia, Ziyang Chen, Xing Wu et al.

Training long-context language models to capture long-range dependencies requires specialized data construction. Current approaches, such as generic text concatenation or heuristic-based variants, frequently fail to guarantee genuine long-range dependencies. We propose EntropyLong, a novel data construction method that leverages predictive uncertainty to verify dependency quality. Our approach identifies high-entropy positions in documents, retrieves semantically relevant contexts from large corpora, and verifies their utility by assessing whether they reduce prediction entropy. This model-in-the-loop verification ensures each dependency represents measurable information gain rather than spurious correlation. We construct training samples with long-range dependencies by combining original documents with these verified contextual supplements. Using FineWebEdu and Cosmopedia, we generate a dataset of 128K-length sequences with verified dependencies. Models trained on this data demonstrate significant improvements on RULER benchmarks, particularly in tasks requiring distant information. Following instruction fine-tuning, our models also achieve substantial gains on LongBenchv2, demonstrating enhanced long-context understanding. Extensive ablation studies further validate the necessity and effectiveness of entropybased verification for long-context training.

Junlong Jia, Xing Wu, Chaochen Gao et al.

High-quality long-context data is essential for training large language models (LLMs) capable of processing extensive documents, yet existing synthesis approaches using relevance-based aggregation face challenges of computational efficiency. We present LiteLong, a resource-efficient method for synthesizing long-context data through structured topic organization and multi-agent debate. Our approach leverages the BISAC book classification system to provide a comprehensive hierarchical topic organization, and then employs a debate mechanism with multiple LLMs to generate diverse, high-quality topics within this structure. For each topic, we use lightweight BM25 retrieval to obtain relevant documents and concatenate them into 128K-token training samples. Experiments on HELMET and Ruler benchmarks demonstrate that LiteLong achieves competitive long-context performance and can seamlessly integrate with other long-dependency enhancement methods. LiteLong makes high-quality long-context data synthesis more accessible by reducing both computational and data engineering costs, facilitating further research in long-context language training.

Duzhen Zhang, Zixiao Wang, Zhong-Zhi Li et al.

The rapid expansion of medical literature presents growing challenges for structuring and integrating domain knowledge at scale. Knowledge Graphs (KGs) offer a promising solution by enabling efficient retrieval, automated reasoning, and knowledge discovery. However, current KG construction methods often rely on supervised pipelines with limited generalizability or naively aggregate outputs from Large Language Models (LLMs), treating biomedical corpora as static and ignoring the temporal dynamics and contextual uncertainty of evolving knowledge. To address these limitations, we introduce MedKGent, a LLM agent framework for constructing temporally evolving medical KGs. Leveraging over 10 million PubMed abstracts published between 1975 and 2023, we simulate the emergence of biomedical knowledge via a fine-grained daily time series. MedKGent incrementally builds the KG in a day-by-day manner using two specialized agents powered by the Qwen2.5-32B-Instruct model. The Extractor Agent identifies knowledge triples and assigns confidence scores via sampling-based estimation, which are used to filter low-confidence extractions and inform downstream processing. The Constructor Agent incrementally integrates the retained triples into a temporally evolving graph, guided by confidence scores and timestamps to reinforce recurring knowledge and resolve conflicts. The resulting KG contains 156,275 entities and 2,971,384 relational triples. Quality assessments by two SOTA LLMs and three domain experts demonstrate an accuracy approaching 90%, with strong inter-rater agreement. To evaluate downstream utility, we conduct RAG across seven medical question answering benchmarks using five leading LLMs, consistently observing significant improvements over non-augmented baselines. Case studies further demonstrate the KG's value in literature-based drug repurposing via confidence-aware causal inference.

Ziyang Chen, Huimu Yu, Xing Wu et al.

Large language models (LLMs) excel in text understanding and generation but raise significant safety and ethical concerns in high-stakes applications. To mitigate these risks, we present Libra-Guard, a cutting-edge safeguard system designed to enhance the safety of Chinese-based LLMs. Leveraging a two-stage curriculum training pipeline, Libra-Guard enhances data efficiency by employing guard pretraining on synthetic samples, followed by fine-tuning on high-quality, real-world data, thereby significantly reducing reliance on manual annotations. To enable rigorous safety evaluations, we also introduce Libra-Test, the first benchmark specifically designed to evaluate the effectiveness of safeguard systems for Chinese content. It covers seven critical harm scenarios and includes over 5,700 samples annotated by domain experts. Experiments show that Libra-Guard achieves 86.79% accuracy, outperforming Qwen2.5-14B-Instruct (74.33%) and ShieldLM-Qwen-14B-Chat (65.69%), and nearing closed-source models like Claude-3.5-Sonnet and GPT-4o. These contributions establish a robust framework for advancing the safety governance of Chinese LLMs and represent a tentative step toward developing safer, more reliable Chinese AI systems.

Yeyong Yu, Xilei Bian, Jie Xiong et al.

With the growing demand for novel materials, machine learning-driven inverse design methods face significant challenges in reconciling the high-dimensional materials composition space with limited experimental data. Existing approaches suffer from two major limitations: (I) machine learning models often lack reliability in high-dimensional spaces, leading to prediction biases during the design process; (II) these models fail to effectively incorporate domain expert knowledge, limiting their capacity to support knowledge-guided inverse design. To address these challenges, we introduce AIMatDesign, a reinforcement learning framework that addresses these limitations by augmenting experimental data using difference-based algorithms to build a trusted experience pool, accelerating model convergence. To enhance model reliability, an automated refinement strategy guided by large language models (LLMs) dynamically corrects prediction inconsistencies, reinforcing alignment between reward signals and state value functions. Additionally, a knowledge-based reward function leverages expert domain rules to improve stability and efficiency during training. Our experiments demonstrate that AIMatDesign significantly surpasses traditional machine learning and reinforcement learning methods in discovery efficiency, convergence speed, and success rates. Among the numerous candidates proposed by AIMatDesign, experimental synthesis of representative Zr-based alloys yielded a top-performing BMG with 1.7GPa yield strength and 10.2\% elongation, closely matching predictions. Moreover, the framework accurately captured the trend of yield strength variation with composition, demonstrating its reliability and potential for closed-loop materials discovery.

Guangyuan Ma, Xing Wu, Zijia Lin et al.

Masked auto-encoder pre-training has emerged as a prevalent technique for initializing and enhancing dense retrieval systems. It generally utilizes additional Transformer decoder blocks to provide sustainable supervision signals and compress contextual information into dense representations. However, the underlying reasons for the effectiveness of such a pre-training technique remain unclear. The usage of additional Transformer-based decoders also incurs significant computational costs. In this study, we aim to shed light on this issue by revealing that masked auto-encoder (MAE) pre-training with enhanced decoding significantly improves the term coverage of input tokens in dense representations, compared to vanilla BERT checkpoints. Building upon this observation, we propose a modification to the traditional MAE by replacing the decoder of a masked auto-encoder with a completely simplified Bag-of-Word prediction task. This modification enables the efficient compression of lexical signals into dense representations through unsupervised pre-training. Remarkably, our proposed method achieves state-of-the-art retrieval performance on several large-scale retrieval benchmarks without requiring any additional parameters, which provides a 67% training speed-up compared to standard masked auto-encoder pre-training with enhanced decoding.

Xing Wu, Chaochen Gao, Meng Lin et al.

Before entering the neural network, a token is generally converted to the corresponding one-hot representation, which is a discrete distribution of the vocabulary. Smoothed representation is the probability of candidate tokens obtained from a pre-trained masked language model, which can be seen as a more informative substitution to the one-hot representation. We propose an efficient data augmentation method, termed text smoothing, by converting a sentence from its one-hot representation to a controllable smoothed representation. We evaluate text smoothing on different benchmarks in a low-resource regime. Experimental results show that text smoothing outperforms various mainstream data augmentation methods by a substantial margin. Moreover, text smoothing can be combined with those data augmentation methods to achieve better performance.

Chaochen Gao, Xing Wu, Peng Wang et al.

Large-scale contrastive learning models can learn very informative sentence embeddings, but are hard to serve online due to the huge model size. Therefore, they often play the role of "teacher", transferring abilities to small "student" models through knowledge distillation. However, knowledge distillation inevitably brings some drop in embedding effect. To tackle that, we propose an effective knowledge distillation framework for contrastive sentence embeddings, termed DistilCSE. It first applies knowledge distillation on a large amount of unlabeled data, and then fine-tunes student models through contrastive learning on limited labeled data. To achieve better distillation results, we further propose Contrastive Knowledge Distillation (CKD). CKD uses InfoNCE as the loss function in knowledge distillation, enhancing the objective consistency among teacher model training, knowledge distillation, and student model fine-tuning. Extensive experiments show that student models trained with the proposed DistilCSE and CKD suffer from little or even no performance decrease and consistently outperform the corresponding counterparts of the same parameter size. Impressively, our 110M student model outperforms the latest state-of-the-art model, i.e., Sentence-T5 (11B), with only 1% parameters and 0.25% unlabeled data.

Xing Wu, Chaochen Gao, Liangjun Zang et al.

Contrastive learning has been attracting much attention for learning unsupervised sentence embeddings. The current state-of-the-art unsupervised method is the unsupervised SimCSE (unsup-SimCSE). Unsup-SimCSE takes dropout as a minimal data augmentation method, and passes the same input sentence to a pre-trained Transformer encoder (with dropout turned on) twice to obtain the two corresponding embeddings to build a positive pair. As the length information of a sentence will generally be encoded into the sentence embeddings due to the usage of position embedding in Transformer, each positive pair in unsup-SimCSE actually contains the same length information. And thus unsup-SimCSE trained with these positive pairs is probably biased, which would tend to consider that sentences of the same or similar length are more similar in semantics. Through statistical observations, we find that unsup-SimCSE does have such a problem. To alleviate it, we apply a simple repetition operation to modify the input sentence, and then pass the input sentence and its modified counterpart to the pre-trained Transformer encoder, respectively, to get the positive pair. Additionally, we draw inspiration from the community of computer vision and introduce a momentum contrast, enlarging the number of negative pairs without additional calculations. The proposed two modifications are applied on positive and negative pairs separately, and build a new sentence embedding method, termed Enhanced Unsup-SimCSE (ESimCSE). We evaluate the proposed ESimCSE on several benchmark datasets w.r.t the semantic text similarity (STS) task. Experimental results show that ESimCSE outperforms the state-of-the-art unsup-SimCSE by an average Spearman correlation of 2.02% on BERT-base.

Xing Wu, Chaochen Gao, Yipeng Su et al.

Contrastive learning has been gradually applied to learn high-quality unsupervised sentence embedding. Among the previous un-supervised methods, the latest state-of-the-art method, as far as we know, is unsupervised SimCSE (unsup-SimCSE). Unsup-SimCSE uses the InfoNCE1loss function in the training stage by pulling semantically similar sentences together and pushing apart dis-similar ones.Theoretically, we expect to use larger batches in unsup-SimCSE to get more adequate comparisons among samples and avoid overfitting. However, increasing the batch size does not always lead to improvements, but instead even lead to performance degradation when the batch size exceeds a threshold. Through statistical observation, we find that this is probably due to the introduction of low-confidence negative pairs after in-creasing the batch size. To alleviate this problem, we introduce a simple smoothing strategy upon the InfoNCE loss function, termedGaussian Smoothing InfoNCE (GS-InfoNCE).Specifically, we add random Gaussian noise vectors as negative samples, which act asa smoothing of the negative sample space.Though being simple, the proposed smooth-ing strategy brings substantial improvements to unsup-SimCSE. We evaluate GS-InfoNCEon the standard semantic text similarity (STS)task. GS-InfoNCE outperforms the state-of-the-art unsup-SimCSE by an average Spear-man correlation of 1.38%, 0.72%, 1.17% and0.28% on the base of BERT-base, BERT-large,RoBERTa-base and RoBERTa-large, respectively.

Xing Wu, Yibing Liu, Xiangyang Zhou et al.

This paper studies compressing pre-trained language models, like BERT (Devlin et al.,2019), via teacher-student knowledge distillation. Previous works usually force the student model to strictly mimic the smoothed labels predicted by the teacher BERT. As an alternative, we propose a new method for BERT distillation, i.e., asking the teacher to generate smoothed word ids, rather than labels, for teaching the student model in knowledge distillation. We call this kind of methodTextSmoothing. Practically, we use the softmax prediction of the Masked Language Model(MLM) in BERT to generate word distributions for given texts and smooth those input texts using that predicted soft word ids. We assume that both the smoothed labels and the smoothed texts can implicitly augment the input corpus, while text smoothing is intuitively more efficient since it can generate more instances in one neural network forward step.Experimental results on GLUE and SQuAD demonstrate that our solution can achieve competitive results compared with existing BERT distillation methods.

Xiaohui Song, Liangjun Zang, Yipeng Su et al.

While several state-of-the-art approaches to dialogue state tracking (DST) have shown promising performances on several benchmarks, there is still a significant performance gap between seen slot values (i.e., values that occur in both training set and test set) and unseen ones (values that occur in training set but not in test set). Recently, the copy-mechanism has been widely used in DST models to handle unseen slot values, which copies slot values from user utterance directly. In this paper, we aim to find out the factors that influence the generalization ability of a common copy-mechanism model for DST. Our key observations include: 1) the copy-mechanism tends to memorize values rather than infer them from contexts, which is the primary reason for unsatisfactory generalization performance; 2) greater diversity of slot values in the training set increase the performance on unseen values but slightly decrease the performance on seen values. Moreover, we propose a simple but effective algorithm of data augmentation to train copy-mechanism models, which augments the input dataset by copying user utterances and replacing the real slot values with randomly generated strings. Users could use two hyper-parameters to realize a trade-off between the performances on seen values and unseen ones, as well as a trade-off between overall performance and computational cost. Experimental results on three widely used datasets (WoZ 2.0, DSTC2, and Multi-WoZ 2.0) show the effectiveness of our approach.

Xing Wu, Dongjun Wei, Liangjun Zang et al.

Structured sentences are important expressions in human writings and dialogues. Previous works on neural text generation fused semantic and structural information by encoding the entire sentence into a mixed hidden representation. However, when a generated sentence becomes complicated, the structure is difficult to be properly maintained. To alleviate this problem, we explicitly separate the modeling process of semantic and structural information. Intuitively, humans generate structured sentences by directly connecting discourses with discourse markers (such as and, but, etc.). Therefore, we propose a task that mimics this process, called discourse transfer. This task represents a structured sentence as (head discourse, discourse marker, tail discourse), and aims at tail discourse generation based on head discourse and discourse marker. We also propose a corresponding model called TransSent, which interprets the relationship between two discourses as a translation1 from the head discourse to the tail discourse in the embedding space. We experiment TransSent not only in discourse transfer task but also in free text generation and dialogue generation tasks. Automatic and human evaluation results show that TransSent can generate structured sentences with high quality, and has certain scalability in different tasks.

Xing Wu, Tao Zhang, Liangjun Zang et al.

This paper focuses on the task of sentiment transfer on non-parallel text, which modifies sentiment attributes (e.g., positive or negative) of sentences while preserving their attribute-independent content. Due to the limited capability of RNNbased encoder-decoder structure to capture deep and long-range dependencies among words, previous works can hardly generate satisfactory sentences from scratch. When humans convert the sentiment attribute of a sentence, a simple but effective approach is to only replace the original sentimental tokens in the sentence with target sentimental expressions, instead of building a new sentence from scratch. Such a process is very similar to the task of Text Infilling or Cloze, which could be handled by a deep bidirectional Masked Language Model (e.g. BERT). So we propose a two step approach "Mask and Infill". In the mask step, we separate style from content by masking the positions of sentimental tokens. In the infill step, we retrofit MLM to Attribute Conditional MLM, to infill the masked positions by predicting words or phrases conditioned on the context1 and target sentiment. We evaluate our model on two review datasets with quantitative, qualitative, and human evaluations. Experimental results demonstrate that our models improve state-of-the-art performance.

Tao Zhang, Xing Wu, Meng Lin et al.

Imbalanced data commonly exists in real world, espacially in sentiment-related corpus, making it difficult to train a classifier to distinguish latent sentiment in text data. We observe that humans often express transitional emotion between two adjacent discourses with discourse markers like "but", "though", "while", etc, and the head discourse and the tail discourse 3 usually indicate opposite emotional tendencies. Based on this observation, we propose a novel plug-and-play method, which first samples discourses according to transitional discourse markers and then validates sentimental polarities with the help of a pretrained attention-based model. Our method increases sample diversity in the first place, can serve as a upstream preprocessing part in data augmentation. We conduct experiments on three public sentiment datasets, with several frequently used algorithms. Results show that our method is found to be consistently effective, even in highly imbalanced scenario, and easily be integrated with oversampling method to boost the performance on imbalanced sentiment classification.

Xing Wu, Shangwen Lv, Liangjun Zang et al.

We propose a novel data augmentation method for labeled sentences called conditional BERT contextual augmentation. Data augmentation methods are often applied to prevent overfitting and improve generalization of deep neural network models. Recently proposed contextual augmentation augments labeled sentences by randomly replacing words with more varied substitutions predicted by language model. BERT demonstrates that a deep bidirectional language model is more powerful than either an unidirectional language model or the shallow concatenation of a forward and backward model. We retrofit BERT to conditional BERT by introducing a new conditional masked language model\footnote{The term "conditional masked language model" appeared once in original BERT paper, which indicates context-conditional, is equivalent to term "masked language model". In our paper, "conditional masked language model" indicates we apply extra label-conditional constraint to the "masked language model".} task. The well trained conditional BERT can be applied to enhance contextual augmentation. Experiments on six various different text classification tasks show that our method can be easily applied to both convolutional or recurrent neural networks classifier to obtain obvious improvement.

Changqing Zou, Haoran Mo, Ruofei Du et al.

We introduce LUCSS, a language-based system for interactive col- orization of scene sketches, based on their semantic understanding. LUCSS is built upon deep neural networks trained via a large-scale repository of scene sketches and cartoon-style color images with text descriptions. It con- sists of three sequential modules. First, given a scene sketch, the segmenta- tion module automatically partitions an input sketch into individual object instances. Next, the captioning module generates the text description with spatial relationships based on the instance-level segmentation results. Fi- nally, the interactive colorization module allows users to edit the caption and produce colored images based on the altered caption. Our experiments show the effectiveness of our approach and the desirability of its compo- nents to alternative choices.