Xin Su

Ankit Pratap Singh, Xin Su, Phillip Howard

Multi-table question answering requires models to retrieve relevant evidence, link schemas, and perform compositional reasoning across relational tables. Existing multi-table Q&A resources typically provide questions and final answers but lack reasoning supervision that explains how answers are derived. To address this gap, we construct a synthetic contrastive reasoning-trace dataset for MMQA by generating validated positive traces and plausible negative traces with heterogeneous LLMs. We then use the resulting preference pairs to fine-tune open-weight LLMs with Contrastive Preference Optimization (CPO). Across Qwen3-14B, Mistral-8B, and Llama-3.1-8B, CPO achieves absolute average improvements over Q&A supervised fine-tuning ranging from 9.7%-16.3%, with gains up to 21 percentage points on MMQA. Ablations show that heterogeneous positive and negative trace generators strengthen the contrastive signal, and automated as well as human evaluations indicate that the generated pairs are largely faithful, coherent, and meaningfully contrastive.

Haonan Guo, Xin Su, Chen Wu et al.

Change detection (CD) is a fundamental and important task for monitoring the land surface dynamics in the earth observation field. Existing deep learning-based CD methods typically extract bi-temporal image features using a weight-sharing Siamese encoder network and identify change regions using a decoder network. These CD methods, however, still perform far from satisfactorily as we observe that 1) deep encoder layers focus on irrelevant background regions and 2) the models' confidence in the change regions is inconsistent at different decoder stages. The first problem is because deep encoder layers cannot effectively learn from imbalanced change categories using the sole output supervision, while the second problem is attributed to the lack of explicit semantic consistency preservation. To address these issues, we design a novel similarity-aware attention flow network (SAAN). SAAN incorporates a similarity-guided attention flow module with deeply supervised similarity optimization to achieve effective change detection. Specifically, we counter the first issue by explicitly guiding deep encoder layers to discover semantic relations from bi-temporal input images using deeply supervised similarity optimization. The extracted features are optimized to be semantically similar in the unchanged regions and dissimilar in the changing regions. The second drawback can be alleviated by the proposed similarity-guided attention flow module, which incorporates similarity-guided attention modules and attention flow mechanisms to guide the model to focus on discriminative channels and regions. We evaluated the effectiveness and generalization ability of the proposed method by conducting experiments on a wide range of CD tasks. The experimental results demonstrate that our method achieves excellent performance on several CD tasks, with discriminative features and semantic consistency preserved.

Xin Su, Dawid Majchrowski, Fangyuan Yu et al.

Large Language Model (LLM) generation remains expensive because autoregressive decoding calls the model once for each new token. Speculative decoding reduces this cost by drafting multiple tokens and verifying them with the target model in one step, but its speedup depends on how many drafted tokens are accepted. Parameter-free draft sources can propose long continuations at low cost in structured and agentic workloads, yet a cache match that looks promising at one generation step may have low payoff at the next. We propose Hybrid Verified Decoding, which predicts the accepted length of a cache draft before verification and uses this payoff estimate to choose between cache verification and a model-based drafter. Across three LLMs and sixteen datasets, Hybrid Verified Decoding is especially effective on agentic workflows, where it outperforms EAGLE3 in every setting with a 2.73x average speedup. Our analysis shows how prompt structure creates cache opportunities, how high-payoff cache drafts concentrate in a small part of the draft space, and how payoff-guided selection reduces sequential decoding work, pointing to runtime draft selection as a promising direction for speculative decoding.

Xin Su, Tiep Le, Steven Bethard et al.

An important open question in the use of large language models for knowledge-intensive tasks is how to effectively integrate knowledge from three sources: the model's parametric memory, external structured knowledge, and external unstructured knowledge. Most existing prompting methods either rely on one or two of these sources, or require repeatedly invoking large language models to generate similar or identical content. In this work, we overcome these limitations by introducing a novel semi-structured prompting approach that seamlessly integrates the model's parametric memory with unstructured knowledge from text documents and structured knowledge from knowledge graphs. Experimental results on open-domain multi-hop question answering datasets demonstrate that our prompting method significantly surpasses existing techniques, even exceeding those that require fine-tuning.

Jingheng Ye, Zishan Xu, Yinghui Li et al.

The paper focuses on the interpretability of Grammatical Error Correction (GEC) evaluation metrics, which received little attention in previous studies. To bridge the gap, we introduce **CLEME2.0**, a reference-based metric describing four fundamental aspects of GEC systems: hit-correction, wrong-correction, under-correction, and over-correction. They collectively contribute to exposing critical qualities and locating drawbacks of GEC systems. Evaluating systems by combining these aspects also leads to superior human consistency over other reference-based and reference-less metrics. Extensive experiments on two human judgment datasets and six reference datasets demonstrate the effectiveness and robustness of our method, achieving a new state-of-the-art result. Our codes are released at https://github.com/THUKElab/CLEME.

Kuo Yang, Zecong Yu, Xin Su et al. · tsinghua

Treatment planning for chronic diseases is a critical task in medical artificial intelligence, particularly in traditional Chinese medicine (TCM). However, generating optimized sequential treatment strategies for patients with chronic diseases in different clinical encounters remains a challenging issue that requires further exploration. In this study, we proposed a TCM herbal prescription planning framework based on deep reinforcement learning for chronic disease treatment (PrescDRL). PrescDRL is a sequential herbal prescription optimization model that focuses on long-term effectiveness rather than achieving maximum reward at every step, thereby ensuring better patient outcomes. We constructed a high-quality benchmark dataset for sequential diagnosis and treatment of diabetes and evaluated PrescDRL against this benchmark. Our results showed that PrescDRL achieved a higher curative effect, with the single-step reward improving by 117% and 153% compared to doctors. Furthermore, PrescDRL outperformed the benchmark in prescription prediction, with precision improving by 40.5% and recall improving by 63%. Overall, our study demonstrates the potential of using artificial intelligence to improve clinical intelligent diagnosis and treatment in TCM.

Xin Su, Phillip Howard, Nagib Hakim et al.

Answering time-sensitive questions from long documents requires temporal reasoning over the times in questions and documents. An important open question is whether large language models can perform such reasoning solely using a provided text document, or whether they can benefit from additional temporal information extracted using other systems. We address this research question by applying existing temporal information extraction systems to construct temporal graphs of events, times, and temporal relations in questions and documents. We then investigate different approaches for fusing these graphs into Transformer models. Experimental results show that our proposed approach for fusing temporal graphs into input text substantially enhances the temporal reasoning capabilities of Transformer models with or without fine-tuning. Additionally, our proposed method outperforms various graph convolution-based approaches and establishes a new state-of-the-art performance on SituatedQA and three splits of TimeQA.

Haonan Guo, Bo Du, Chen Wu et al.

The efficacy of building footprint segmentation from remotely sensed images has been hindered by model transfer effectiveness. Many existing building segmentation methods were developed upon the encoder-decoder architecture of U-Net, in which the encoder is finetuned from the newly developed backbone networks that are pre-trained on ImageNet. However, the heavy computational burden of the existing decoder designs hampers the successful transfer of these modern encoder networks to remote sensing tasks. Even the widely-adopted deep supervision strategy fails to mitigate these challenges due to its invalid loss in hybrid regions where foreground and background pixels are intermixed. In this paper, we conduct a comprehensive evaluation of existing decoder network designs for building footprint segmentation and propose an efficient framework denoted as BFSeg to enhance learning efficiency and effectiveness. Specifically, a densely-connected coarse-to-fine feature fusion decoder network that facilitates easy and fast feature fusion across scales is proposed. Moreover, considering the invalidity of hybrid regions in the down-sampled ground truth during the deep supervision process, we present a lenient deep supervision and distillation strategy that enables the network to learn proper knowledge from deep supervision. Building upon these advancements, we have developed a new family of building segmentation networks, which consistently surpass prior works with outstanding performance and efficiency across a wide range of newly developed encoder networks.

Han Nie, Bin Luo, Jun Liu et al.

We present REMM, a rotation-equivariant framework for end-to-end multimodal image matching, which fully encodes rotational differences of descriptors in the whole matching pipeline. Previous learning-based methods mainly focus on extracting modal-invariant descriptors, while consistently ignoring the rotational invariance. In this paper, we demonstrate that our REMM is very useful for multimodal image matching, including multimodal feature learning module and cyclic shift module. We first learn modal-invariant features through the multimodal feature learning module. Then, we design the cyclic shift module to rotationally encode the descriptors, greatly improving the performance of rotation-equivariant matching, which makes them robust to any angle. To validate our method, we establish a comprehensive rotation and scale-matching benchmark for evaluating the anti-rotation performance of multimodal images, which contains a combination of multi-angle and multi-scale transformations from four publicly available datasets. Extensive experiments show that our method outperforms existing methods in benchmarking and generalizes well to independent datasets. Additionally, we conducted an in-depth analysis of the key components of the REMM to validate the improvements brought about by the cyclic shift module. Code and dataset at https://github.com/HanNieWHU/REMM.

Haonan Guo, Xin Su, Chen Wu et al.

Buildings are the basic carrier of social production and human life; roads are the links that interconnect social networks. Building and road information has important application value in the frontier fields of regional coordinated development, disaster prevention, auto-driving, etc. Mapping buildings and roads from very high-resolution (VHR) remote sensing images have become a hot research topic. However, the existing methods often ignore the strong spatial correlation between roads and buildings and extract them in isolation. To fully utilize the complementary advantages between buildings and roads, we propose a building-road collaborative extraction method based on multi-task and cross-scale feature interaction to improve the accuracy of both tasks in a complementary way. A multi-task interaction module is proposed to interact information across tasks and preserve the unique information of each task, which tackle the seesaw phenomenon in multitask learning. By considering the variation in appearance and structure between buildings and roads, a cross-scale interaction module is designed to automatically learn the optimal reception field for different tasks. Compared with many existing methods that train each task individually, the proposed collaborative extraction method can utilize the complementary advantages between buildings and roads by the proposed inter-task and inter-scale feature interactions, and automatically select the optimal reception field for different tasks. Experiments on a wide range of urban and rural scenarios show that the proposed algorithm can achieve building-road extraction with outstanding performance and efficiency.

Xiaolei Qin, Xin Su, Liangpei Zhang

Satellite image time series (SITS) data provides continuous observations over time, allowing for the tracking of vegetation changes and growth patterns throughout the seasons and years. Numerous deep learning (DL) approaches using SITS for crop classification have emerged recently, with the latest approaches adopting Transformer for SITS classification. However, the quadratic complexity of self-attention in Transformer poses challenges for classifying long time series. While the cutting-edge Mamba architecture has demonstrated strength in various domains, including remote sensing image interpretation, its capacity to learn temporal representations in SITS data remains unexplored. Moreover, the existing SITS classification methods often depend solely on crop labels as supervision signals, which fails to fully exploit the temporal information. In this paper, we proposed a Satellite Image Time Series Mamba (SITSMamba) method for crop classification based on remote sensing time series data. The proposed SITSMamba contains a spatial encoder based on Convolutional Neural Networks (CNN) and a Mamba-based temporal encoder. To exploit richer temporal information from SITS, we design two branches of decoder used for different tasks. The first branch is a crop Classification Branch (CBranch), which includes a ConvBlock to decode the feature to a crop map. The second branch is a SITS Reconstruction Branch that uses a Linear layer to transform the encoded feature to predict the original input values. Furthermore, we design a Positional Weight (PW) applied to the RBranch to help the model learn rich latent knowledge from SITS. We also design two weighting factors to control the balance of the two branches during training. The code of SITSMamba is available at: https://github.com/XiaoleiQinn/SITSMamba.

Haonan Guo, Xin Su, Chen Wu et al.

Recently, the flourishing large language models(LLM), especially ChatGPT, have shown exceptional performance in language understanding, reasoning, and interaction, attracting users and researchers from multiple fields and domains. Although LLMs have shown great capacity to perform human-like task accomplishment in natural language and natural image, their potential in handling remote sensing interpretation tasks has not yet been fully explored. Moreover, the lack of automation in remote sensing task planning hinders the accessibility of remote sensing interpretation techniques, especially to non-remote sensing experts from multiple research fields. To this end, we present Remote Sensing ChatGPT, an LLM-powered agent that utilizes ChatGPT to connect various AI-based remote sensing models to solve complicated interpretation tasks. More specifically, given a user request and a remote sensing image, we utilized ChatGPT to understand user requests, perform task planning according to the tasks' functions, execute each subtask iteratively, and generate the final response according to the output of each subtask. Considering that LLM is trained with natural language and is not capable of directly perceiving visual concepts as contained in remote sensing images, we designed visual cues that inject visual information into ChatGPT. With Remote Sensing ChatGPT, users can simply send a remote sensing image with the corresponding request, and get the interpretation results as well as language feedback from Remote Sensing ChatGPT. Experiments and examples show that Remote Sensing ChatGPT can tackle a wide range of remote sensing tasks and can be extended to more tasks with more sophisticated models such as the remote sensing foundation model. The code and demo of Remote Sensing ChatGPT is publicly available at https://github.com/HaonanGuo/Remote-Sensing-ChatGPT .

Xudong Yan, Songhe Feng, Jiaxin Wang et al.

Compositional Zero-Shot Learning (CZSL) aims to recognize novel attribute-object compositions based on the knowledge learned from seen ones. Existing methods suffer from performance degradation caused by the distribution shift of label space at test time, which stems from the inclusion of unseen compositions recombined from attributes and objects. To overcome the challenge, we propose a novel approach that accumulates comprehensive knowledge in both textual and visual modalities from unsupervised data to update multimodal prototypes at test time. Building on this, we further design an adaptive update weight to control the degree of prototype adjustment, enabling the model to flexibly adapt to distribution shift during testing. Moreover, a dynamic priority queue is introduced that stores high-confidence images to acquire visual prototypes from historical images for inference. Since the model tends to favor compositions already stored in the queue during testing, we warm-start the queue by initializing it with training images for visual prototypes of seen compositions and generating unseen visual prototypes using the mapping learned between seen and unseen textual prototypes. Considering the semantic consistency of multimodal knowledge, we align textual and visual prototypes by multimodal collaborative representation learning. To provide a more reliable evaluation for CZSL, we introduce a new benchmark dataset, C-Fashion, and refine the widely used but noisy MIT-States dataset. Extensive experiments indicate that our approach achieves state-of-the-art performance on four benchmark datasets under both closed-world and open-world settings. The source code and datasets are available at https://github.com/xud-yan/WARM-CAT .

Weixing Liu, Jun Liu, Xin Su et al.

Recent studies have used unsupervised domain adaptive object detection (UDAOD) methods to bridge the domain gap in remote sensing (RS) images. However, UDAOD methods typically assume that the source domain data can be accessed during the domain adaptation process. This setting is often impractical in the real world due to RS data privacy and transmission difficulty. To address this challenge, we propose a practical source-free object detection (SFOD) setting for RS images, which aims to perform target domain adaptation using only the source pre-trained model. We propose a new SFOD method for RS images consisting of two parts: perturbed domain generation and alignment. The proposed multilevel perturbation constructs the perturbed domain in a simple yet efficient form by perturbing the domain-variant features at the image level and feature level according to the color and style bias. The proposed multilevel alignment calculates feature and label consistency between the perturbed domain and the target domain across the teacher-student network, and introduces the distillation of feature prototype to mitigate the noise of pseudo-labels. By requiring the detector to be consistent in the perturbed domain and the target domain, the detector is forced to focus on domaininvariant features. Extensive results of three synthetic-to-real experiments and three cross-sensor experiments have validated the effectiveness of our method which does not require access to source domain RS images. Furthermore, experiments on computer vision datasets show that our method can be extended to other fields as well. Our code will be available at: https://weixliu.github.io/ .

Xin Su, Zhuoran Zheng, Chen Wu

All-in-one image restoration tasks are becoming increasingly important, especially for ultra-high-definition (UHD) images. Existing all-in-one UHD image restoration methods usually boost the model's performance by introducing prompt or customized dynamized networks for different degradation types. For the inference stage, it might be friendly, but in the training stage, since the model encounters multiple degraded images of different quality in an epoch, these cluttered learning objectives might be information pollution for the model. To address this problem, we propose a new training paradigm for general image restoration models, which we name \textbf{Review Learning}, which enables image restoration models to be capable enough to handle multiple types of degradation without prior knowledge and prompts. This approach begins with sequential training of an image restoration model on several degraded datasets, combined with a review mechanism that enhances the image restoration model's memory for several previous classes of degraded datasets. In addition, we design a lightweight all-purpose image restoration network that can efficiently reason about degraded images with 4K ($3840 \times 2160$) resolution on a single consumer-grade GPU.

Nirmalendu Prakash, Narmeen Fatimah Oozeer, Xin Su et al.

CLIP retrieval is typically framed as a pointwise similarity problem in a shared embedding space. While CLIP achieves strong global cross-modal alignment, many retrieval failures arise from local geometric inconsistencies: nearby items are incorrectly ordered, leading to systematic confusions (e.g., pentagon vs. hexagon) and produces diffuse, weakly controlled result sets. Prior work largely optimizes for point wise relevance or finetuning to mitigate these problems. We instead view retrieval as a problem of neighborhood alignment. Our work introduces (1) neighborhood-level re-ranking via Hungarian matching, which rewards structural consistency; (2) query-conditioned local steering, where directions derived from contrastive neighborhoods around the query reshape retrieval. We show that these techniques improve retrieval performance on attribute-binding and compositional retrieval tasks. Together, these methods operate on local neighborhoods but serve different roles: re-ranking rewards alignment whereas local steering controls neighborhood structure. This shows that retrieval quality and controllability depend critically on local structure, which can be exploited at inference time without retraining.

Xin Su, Yao Zhou, Zifei Shan et al.

It is a long-standing challenge in modern recommender systems to effectively make recommendations for new users, namely the cold-start problem. Cross-Domain Recommendation (CDR) has been proposed to address this challenge, but current ways to represent users' interests across systems are still severely limited. We introduce Personal Knowledge Graph (PKG) as a domain-invariant interest representation, and propose a novel CDR paradigm named MeKB-Rec. We first link users and entities in a knowledge base to construct a PKG of users' interests, named MeKB. Then we learn a semantic representation of MeKB for the cross-domain recommendation. To efficiently utilize limited training data in CDR, MeKB-Rec employs Pretrained Language Models to inject world knowledge into understanding users' interests. Beyond most existing systems, our approach builds a semantic mapping across domains which breaks the requirement for in-domain user behaviors, enabling zero-shot recommendations for new users in a low-resource domain. We experiment MeKB-Rec on well-established public CDR datasets, and demonstrate that the new formulation % is more powerful than previous approaches, achieves a new state-of-the-art that significantly improves HR@10 and NDCG@10 metrics over best previous approaches by 24\%--91\%, with a 105\% improvement for HR@10 of zero-shot users with no behavior in the target domain. We deploy MeKB-Rec in WeiXin recommendation scenarios and achieve significant gains in core online metrics. MeKB-Rec is now serving hundreds of millions of users in real-world products.

Fangyuan Yu, Xin Su, Amir Abdullah

We investigate the temporal concatenation of sub-policies in Markov Decision Processes (MDP) with time-varying reward functions. We introduce General Dijkstra Search (GDS), and prove that globally optimal goal-reaching policies can be recovered through temporal composition of intermediate optimal sub-policies. Motivated by the "search, select, update" principle underlying GDS, we propose Dynamic Latent Routing (DLR), a language-model post-training method that jointly learns discrete latent codes, routing policies, and model parameters through dynamic search in a single training stage. In low-data fine-tuning settings, DLR matches or outperforms supervised fine-tuning across four datasets and six models, achieving a mean gain of +6.6 percentage points, while prior discrete-latent baselines consistently underperform SFT. Mechanistic analyses and targeted code ablations show that DLR learns structured routing behaviors with distinct causal roles.

Phillip Howard, Xin Su, Kathleen C. Fraser

The rapid adoption of large vision-language models (LVLMs) in recent years has been accompanied by growing fairness concerns due to their propensity to reinforce harmful societal stereotypes. While significant attention has been paid to such fairness concerns in the context of social biases, relatively little prior work has examined the presence of stereotypes in LVLMs related to cultural contexts such as religion, nationality, and socioeconomic status. In this work, we aim to narrow this gap by investigating how cultural contexts depicted in images influence the judgments LVLMs make about a person's moral, ethical, and political values. We conduct a multi-dimensional analysis of such value judgments in five popular LVLMs using counterfactual image sets, which depict the same person across different cultural contexts. Our evaluation framework diagnoses LVLM awareness of cultural value differences through the use of Moral Foundations Theory, lexical analyses, and the sensitivity of generated values to depicted cultural contexts.

Phillip Howard, Xin Su, Kathleen C. Fraser

Large Vision-Language Models (LVLMs) have grown increasingly powerful in recent years, but can also exhibit harmful biases. Prior studies investigating such biases have primarily focused on demographic traits related to the visual characteristics of a person depicted in an image, such as their race or gender. This has left biases related to cultural differences (e.g., religion, socioeconomic status), which cannot be readily discerned from an individual's appearance alone, relatively understudied. A key challenge in measuring cultural biases is that determining which group an individual belongs to often depends upon cultural context cues in images, and datasets annotated with cultural context cues are lacking. To address this gap, we introduce Cultural Counterfactuals: a high-quality synthetic dataset containing nearly 60k counterfactual images for measuring cultural biases related to religion, nationality, and socioeconomic status. To ensure that cultural contexts are accurately depicted, we generate our dataset using an image-editing model to place people of different demographics into real cultural context images. This enables the construction of counterfactual image sets which depict the same person in multiple different contexts, allowing for precise measurement of the impact that cultural context differences have on LVLM outputs. We demonstrate the utility of Cultural Counterfactuals for quantifying cultural biases in popular LVLMs.

Xin Su, Zhuoran Zheng

Privacy protection has always been an ongoing topic, especially for AI. Currently, a low-cost scheme called Machine Unlearning forgets the private data remembered in the model. Specifically, given a private dataset and a trained neural network, we need to use e.g. pruning, fine-tuning, and gradient ascent to remove the influence of the private dataset on the neural network. Inspired by this, we try to use this concept to bridge the gap between the fields of image restoration and security, creating a new research idea. We propose the scene for the All-In-One model (a neural network that restores a wide range of degraded information), where a given dataset such as haze, or rain, is private and needs to be eliminated from the influence of it on the trained model. Notably, we find great challenges in this task to remove the influence of sensitive data while ensuring that the overall model performance remains robust, which is akin to directing a symphony orchestra without specific instruments while keeping the playing soothing. Here we explore a simple but effective approach: Instance-wise Unlearning through the use of adversarial examples and gradient ascent techniques. Our approach is a low-cost solution compared to the strategy of retraining the model from scratch, where the gradient ascent trick forgets the specified data and the performance of the adversarial sample maintenance model is robust. Through extensive experimentation on two popular unified image restoration models, we show that our approach effectively preserves knowledge of remaining data while unlearning a given degradation type.

Xiaolei Qin, Di Wang, Jing Zhang et al.

Satellite image time series (SITS) provide continuous observations of the Earth's surface, making them essential for applications such as environmental management and disaster assessment. However, existing spatiotemporal foundation models rely on plain vision transformers, which encode entire temporal sequences without explicitly capturing multiscale spatiotemporal relationships between land objects. This limitation hinders their effectiveness in downstream tasks. To overcome this challenge, we propose TiMo, a novel hierarchical vision transformer foundation model tailored for SITS analysis. At its core, we introduce a spatiotemporal gyroscope attention mechanism that dynamically captures evolving multiscale patterns across both time and space. For pre-training, we curate MillionST, a large-scale dataset of one million images from 100,000 geographic locations, each captured across 10 temporal phases over five years, encompassing diverse geospatial changes and seasonal variations. Leveraging this dataset, we adapt masked image modeling to pre-train TiMo, enabling it to effectively learn and encode generalizable spatiotemporal representations.Extensive experiments across multiple spatiotemporal tasks-including deforestation monitoring, land cover segmentation, crop type classification, and flood detection-demonstrate TiMo's superiority over state-of-the-art methods. Code, model, and dataset will be released at https://github.com/MiliLab/TiMo.

Prafulla Kumar Choubey, Xin Su, Man Luo et al.

Knowledge graphs (KGs) generated by large language models (LLMs) are becoming increasingly valuable for Retrieval-Augmented Generation (RAG) applications that require knowledge-intensive reasoning. However, existing KG extraction methods predominantly rely on prompt-based approaches, which are inefficient for processing large-scale corpora. These approaches often suffer from information loss, particularly with long documents, due to the lack of specialized design for KG construction. Additionally, there is a gap in evaluation datasets and methodologies for ontology-free KG construction. To overcome these limitations, we propose SynthKG, a multi-step, document-level ontology-free KG synthesis workflow based on LLMs. By fine-tuning a smaller LLM on the synthesized document-KG pairs, we streamline the multi-step process into a single-step KG generation approach called Distill-SynthKG, substantially reducing the number of LLM inference calls. Furthermore, we re-purpose existing question-answering datasets to establish KG evaluation datasets and introduce new evaluation metrics. Using KGs produced by Distill-SynthKG, we also design a novel graph-based retrieval framework for RAG. Experimental results demonstrate that Distill-SynthKG not only surpasses all baseline models in KG quality -- including models up to eight times larger -- but also consistently excels in retrieval and question-answering tasks. Our proposed graph retrieval framework also outperforms all KG-retrieval methods across multiple benchmark datasets. We release the SynthKG dataset and Distill-SynthKG model publicly to support further research and development.

Neale Ratzlaff, Man Luo, Xin Su et al.

Multimodal models typically combine a powerful large language model (LLM) with a vision encoder and are then trained on multimodal data via instruction tuning. While this process adapts LLMs to multimodal settings, it remains unclear whether this adaptation compromises their original language reasoning capabilities. In this work, we explore the effects of multimodal instruction tuning on language reasoning performance. We focus on LLaVA, a leading multimodal framework that integrates LLMs such as Vicuna or Mistral with the CLIP vision encoder. We compare the performance of the original LLMs with their multimodal-adapted counterparts across eight language reasoning tasks. Our experiments yield several key insights. First, the impact of multimodal learning varies between Vicuna and Mistral: we observe a degradation in language reasoning for Mistral but improvements for Vicuna across most tasks. Second, while multimodal instruction learning consistently degrades performance on mathematical reasoning tasks (e.g., GSM8K), it enhances performance on commonsense reasoning tasks (e.g., CommonsenseQA). Finally, we demonstrate that a training-free model merging technique can effectively mitigate the language reasoning degradation observed in multimodal-adapted Mistral and even improve performance on visual tasks.

Shaoshen Chen, Yangning Li, Zishan Xu et al.

Large Language Models (LLMs) face computational inefficiencies and redundant processing when handling long context inputs, prompting a focus on compression techniques. While existing semantic vector-based compression methods achieve promising performance, these methods fail to account for the intrinsic information density variations between context chunks, instead allocating soft tokens uniformly across context chunks. This uniform distribution inevitably diminishes allocation to information-critical regions. To address this, we propose Dynamic Allocation of Soft Tokens (DAST), a simple yet effective method that leverages the LLM's intrinsic understanding of contextual relevance to guide compression. DAST combines perplexity-based local information with attention-driven global information to dynamically allocate soft tokens to the informative-rich chunks, enabling effective, context-aware compression. Experimental results across multiple benchmarks demonstrate that DAST surpasses state-of-the-art methods.

Sezen Perçin, Xin Su, Qutub Sha Syed et al.

Large language models (LLMs) are very costly and inefficient to update with new information. To address this limitation, retrieval-augmented generation (RAG) has been proposed as a solution that dynamically incorporates external knowledge during inference, improving factual consistency and reducing hallucinations. Despite its promise, RAG systems face practical challenges-most notably, a strong dependence on the quality of the input query for accurate retrieval. In this paper, we investigate the sensitivity of different components in the RAG pipeline to various types of query perturbations. Our analysis reveals that the performance of commonly used retrievers can degrade significantly even under minor query variations. We study each module in isolation as well as their combined effect in an end-to-end question answering setting, using both general-domain and domain-specific datasets. Additionally, we propose an evaluation framework to systematically assess the query-level robustness of RAG pipelines and offer actionable recommendations for practitioners based on the results of more than 1092 experiments we performed.

Chenguang Liu, Chisheng Wang, Feifei Dong et al.

Edge detection is a long-standing problem in computer vision. Despite the efficiency of existing algorithms, their performance, however, rely heavily on the pre-trained weights of the backbone network on the ImageNet dataset. The use of pre-trained weights in previous methods significantly increases the difficulty to design new models for edge detection without relying on existing well-trained ImageNet models, as pre-training the model on the ImageNet dataset is expensive and becomes compulsory to ensure the fairness of comparison. Besides, the pre-training and fine-tuning strategy is not always useful and sometimes even inaccessible. For instance, the pre-trained weights on the ImageNet dataset are unlikely to be helpful for edge detection in Synthetic Aperture Radar (SAR) images due to strong differences in the statistics between optical images and SAR images. Moreover, no dataset has comparable size to the ImageNet dataset for SAR image processing. In this work, we study the performance achievable by state-of-the-art deep learning based edge detectors in publicly available datasets when they are trained from scratch, and devise a new network architecture, the multi-stream and multi-scale fusion net (msmsfnet), for edge detection. We show in our experiments that by training all models from scratch, our model outperforms state-of-the-art edge detectors in three publicly available datasets. We also demonstrate the efficiency of our model for edge detection in SAR images, where no useful pre-trained weight is available. Finally, We show that our model is able to achieve competitive performance on the BSDS500 dataset when the pre-trained weights are used.

Tingwei Lu, Yangning Li, Liyuan Wang et al.

The emergence of large language models (LLMs) has significantly promoted the development of code generation task, sparking a surge in pertinent literature. Current research is hindered by redundant generation results and a tendency to overfit local patterns in the short term. Although existing studies attempt to alleviate the issue by adopting a multi-token prediction strategy, there remains limited focus on choosing the appropriate processing length for generations. By analyzing the attention between tokens during the generation process of LLMs, it can be observed that the high spikes of the attention scores typically appear at the end of lines. This insight suggests that it is reasonable to treat each line of code as a fundamental processing unit and generate them sequentially. Inspired by this, we propose the \textbf{LSR-MCTS} algorithm, which leverages MCTS to determine the code line-by-line and select the optimal path. Further, we integrate a self-refine mechanism at each node to enhance diversity and generate higher-quality programs through error correction. Extensive experiments and comprehensive analyses on three public coding benchmarks demonstrate that our method outperforms the state-of-the-art performance approaches.

Man Luo, David Cobbley, Xin Su et al.

Computer use agents (CUA) are systems that automatically interact with graphical user interfaces (GUIs) to complete tasks. CUA have made significant progress with the advent of large vision-language models (VLMs). However, these agents typically rely on cloud-based inference with substantial compute demands, raising critical privacy and scalability concerns, especially when operating on personal devices. In this work, we take a step toward privacy-preserving and resource-efficient agents by developing a lightweight vision-language model that runs entirely on local machines. To train this compact agent, we introduce an LLM-as-Judge framework that automatically evaluates and filters synthetic interaction trajectories, producing high-quality data for reinforcement learning without human annotation. Experiments on the OS-World benchmark demonstrate that our fine-tuned local model outperforms existing baselines, highlighting a promising path toward private, efficient, and generalizable GUI agents.

Zhuoran Zheng, Xin Su, Chen Wu et al.

With the exponential increase in image data, training an image restoration model is laborious. Dataset distillation is a potential solution to this problem, yet current distillation techniques are a blank canvas in the field of image restoration. To fill this gap, we propose the Distribution-aware Dataset Distillation method (TripleD), a new framework that extends the principles of dataset distillation to image restoration. Specifically, TripleD uses a pre-trained vision Transformer to extract features from images for complexity evaluation, and the subset (the number of samples is much smaller than the original training set) is selected based on complexity. The selected subset is then fed through a lightweight CNN that fine-tunes the image distribution to align with the distribution of the original dataset at the feature level. To efficiently condense knowledge, the training is divided into two stages. Early stages focus on simpler, low-complexity samples to build foundational knowledge, while later stages select more complex and uncertain samples as the model matures. Our method achieves promising performance on multiple image restoration tasks, including multi-task image restoration, all-in-one image restoration, and ultra-high-definition image restoration tasks. Note that we can train a state-of-the-art image restoration model on an ultra-high-definition (4K resolution) dataset using only one consumer-grade GPU in less than 8 hours (500 savings in computing resources and immeasurable training time).

Xin Su, Chen Wu, Zhuoran Zheng

Models of dense prediction based on traditional Artificial Neural Networks (ANNs) require a lot of energy, especially for image restoration tasks. Currently, neural networks based on the SNN (Spiking Neural Network) framework are beginning to make their mark in the field of image restoration, especially as they typically use less than 10\% of the energy of ANNs with the same architecture. However, training an SNN is much more expensive than training an ANN, due to the use of the heuristic gradient descent strategy. In other words, the process of SNN's potential membrane signal changing from sparse to dense is very slow, which affects the convergence of the whole model.To tackle this problem, we propose a novel distillation technique, called asymmetric framework (ANN-SNN) distillation, in which the teacher is an ANN and the student is an SNN. Specifically, we leverage the intermediate features (feature maps) learned by the ANN as hints to guide the training process of the SNN. This approach not only accelerates the convergence of the SNN but also improves its final performance, effectively bridging the gap between the efficiency of the SNN and the superior learning capabilities of ANN. Extensive experimental results show that our designed SNN-based image restoration model, which has only 1/300 the number of parameters of the teacher network and 1/50 the energy consumption of the teacher network, is as good as the teacher network in some denoising tasks.

Xin Su, Phillip Howard, Steven Bethard

Temporal information extraction (IE) aims to extract structured temporal information from unstructured text, thereby uncovering the implicit timelines within. This technique is applied across domains such as healthcare, newswire, and intelligence analysis, aiding models in these areas to perform temporal reasoning and enabling human users to grasp the temporal structure of text. Transformer-based pre-trained language models have produced revolutionary advancements in natural language processing, demonstrating exceptional performance across a multitude of tasks. Despite the achievements garnered by Transformer-based approaches in temporal IE, there is a lack of comprehensive reviews on these endeavors. In this paper, we aim to bridge this gap by systematically summarizing and analyzing the body of work on temporal IE using Transformers while highlighting potential future research directions.

Ding Zhang, Yangning Li, Lichen Bai et al.

Chinese grammatical error correction (CGEC) aims to detect and correct errors in the input Chinese sentences. Recently, Pre-trained Language Models (PLMS) have been employed to improve the performance. However, current approaches ignore that correction difficulty varies across different instances and treat these samples equally, enhancing the challenge of model learning. To address this problem, we propose a multi-granularity Curriculum Learning (CL) framework. Specifically, we first calculate the correction difficulty of these samples and feed them into the model from easy to hard batch by batch. Then Instance-Level CL is employed to help the model optimize in the appropriate direction automatically by regulating the loss function. Extensive experimental results and comprehensive analyses of various datasets prove the effectiveness of our method.

Xin Su, Chen Wu, Zhuoran Zheng

Ultra-high-definition (UHD) image restoration is vital for applications demanding exceptional visual fidelity, yet existing methods often face a trade-off between restoration quality and efficiency, limiting their practical deployment. In this paper, we propose TSFormer, an all-in-one framework that integrates \textbf{T}rusted learning with \textbf{S}parsification to boost both generalization capability and computational efficiency in UHD image restoration. The key is that only a small amount of token movement is allowed within the model. To efficiently filter tokens, we use Min-$p$ with random matrix theory to quantify the uncertainty of tokens, thereby improving the robustness of the model. Our model can run a 4K image in real time (40fps) with 3.38 M parameters. Extensive experiments demonstrate that TSFormer achieves state-of-the-art restoration quality while enhancing generalization and reducing computational demands. In addition, our token filtering method can be applied to other image restoration models to effectively accelerate inference and maintain performance.

Jia Wei, Xin Su, Chun Ouyang

Integrating Internet of Things (IoT) data with business process event logs is crucial for analysing IoT-enhanced processes, yet remains challenging due to differences in abstraction levels and the separation of data sources. Simply incorporating raw IoT data increases the size and complexity of the resulting log, often requiring additional processing before process analysis can be performed. While tools for generating IoT-enriched event logs exist, they either rely on specialised schemas or focus on extracting event logs from sensor data, offering limited support for integrating process-relevant IoT data into existing event logs. To address this gap, we present IOTEL, a tool for systematically generating IoT-enriched object-centric event logs (OCEL). By building on the OCEL schema, IOTEL enables structured IoT data integration compatible with existing process mining tools. It support practitioners and researchers in analysing IoT-enhanced business processes, as demonstrated in a real-world scenario. A video demonstrating the tool is available online.

Xin Su, Sungduk Yu, Phillip Howard et al.

Time normalization is the task of converting natural language temporal expressions into machine-readable representations. It underpins many downstream applications in information retrieval, question answering, and clinical decision-making. Traditional systems based on the ISO-TimeML schema limit expressivity and struggle with complex constructs such as compositional, event-relative, and multi-span time expressions. In this work, we introduce a novel formulation of time normalization as a code generation task grounded in the SCATE framework, which defines temporal semantics through symbolic and compositional operators. We implement a fully executable SCATE Python library and demonstrate that large language models (LLMs) can generate executable SCATE code. Leveraging this capability, we develop an automatic data augmentation pipeline using LLMs to synthesize large-scale annotated data with code-level validation. Our experiments show that small, locally deployable models trained on this augmented data can achieve strong performance, outperforming even their LLM parents and enabling practical, accurate, and interpretable time normalization.

Xin Su, Chen Wu, Yu Zhang et al.

Restoring images afflicted by complex real-world degradations remains challenging, as conventional methods often fail to adapt to the unique mixture and severity of artifacts present. This stems from a reliance on indirect cues which poorly capture the true perceptual quality deficit. To address this fundamental limitation, we introduce AdaQual-Diff, a diffusion-based framework that integrates perceptual quality assessment directly into the generative restoration process. Our approach establishes a mathematical relationship between regional quality scores from DeQAScore and optimal guidance complexity, implemented through an Adaptive Quality Prompting mechanism. This mechanism systematically modulates prompt structure according to measured degradation severity: regions with lower perceptual quality receive computationally intensive, structurally complex prompts with precise restoration directives, while higher quality regions receive minimal prompts focused on preservation rather than intervention. The technical core of our method lies in the dynamic allocation of computational resources proportional to degradation severity, creating a spatially-varying guidance field that directs the diffusion process with mathematical precision. By combining this quality-guided approach with content-specific conditioning, our framework achieves fine-grained control over regional restoration intensity without requiring additional parameters or inference iterations. Experimental results demonstrate that AdaQual-Diff achieves visually superior restorations across diverse synthetic and real-world datasets.

Xin Su, Zhuoran Zheng

With the rising imaging resolution of handheld devices, existing multi-exposure image fusion algorithms struggle to generate a high dynamic range image with ultra-high resolution in real-time. Apart from that, there is a trend to design a manageable and editable algorithm as the different needs of real application scenarios. To tackle these issues, we introduce 3D LUT technology, which can enhance images with ultra-high-definition (UHD) resolution in real time on resource-constrained devices. However, since the fusion of information from multiple images with different exposure rates is uncertain, and this uncertainty significantly trials the generalization power of the 3D LUT grid. To address this issue and ensure a robust learning space for the model, we propose using a teacher-student network to model the uncertainty on the 3D LUT grid.Furthermore, we provide an editable mode for the multi-exposure image fusion algorithm by using the implicit representation function to match the requirements in different scenarios. Extensive experiments demonstrate that our proposed method is highly competitive in efficiency and accuracy.

Yi Xiao, Bin Luo, Jun Liu et al.

Change Detection (CD) enables the identification of alterations between images of the same area captured at different times. However, existing CD methods still struggle to address pseudo changes resulting from domain information differences in multi-temporal images and instances of detail errors caused by the loss and contamination of detail features during the upsampling process in the network. To address this, we propose a bi-temporal Gaussian distribution feature-dependent network (BGFD). Specifically, we first introduce the Gaussian noise domain disturbance (GNDD) module, which approximates distribution using image statistical features to characterize domain information, samples noise to perturb the network for learning redundant domain information, addressing domain information differences from a more fundamental perspective. Additionally, within the feature dependency facilitation (FDF) module, we integrate a novel mutual information difference loss ($L_{MI}$) and more sophisticated attention mechanisms to enhance the capabilities of the network, ensuring the acquisition of essential domain information. Subsequently, we have designed a novel detail feature compensation (DFC) module, which compensates for detail feature loss and contamination introduced during the upsampling process from the perspectives of enhancing local features and refining global features. The BGFD has effectively reduced pseudo changes and enhanced the detection capability of detail information. It has also achieved state-of-the-art performance on four publicly available datasets - DSIFN-CD, SYSU-CD, LEVIR-CD, and S2Looking, surpassing baseline models by +8.58%, +1.28%, +0.31%, and +3.76% respectively, in terms of the F1-Score metric.

Xin Su, Man Luo, Kris W Pan et al.

Multimodal retrieval augmented generation (RAG) plays a crucial role in domains such as knowledge-based visual question answering (KB-VQA), where external knowledge is needed to answer a question. However, existing multimodal LLMs (MLLMs) are not designed for context-augmented generation, limiting their effectiveness in such tasks. While synthetic data generation has recently gained attention for training MLLMs, its application for context-augmented generation remains underexplored. To address this gap, we introduce SK-VQA, a large-scale synthetic multimodal dataset containing over 2 million visual question-answer pairs, each associated with context documents containing information necessary to determine the final answer. Compared to previous datasets, SK-VQA contains 11x more unique questions, exhibits greater domain diversity, and covers a broader spectrum of image sources. Through human evaluations, we confirm the high quality of the generated question-answer pairs and their contextual relevance. Extensive experiments show that SK-VQA serves both as a challenging KB-VQA benchmark and as an effective training resource for adapting MLLMs to context-augmented generation. Our results further indicate that models trained on SK-VQA demonstrate enhanced generalization in both context-aware VQA and multimodal RAG settings. SK-VQA is publicly available via Hugging Face Hub.

Xin Su, Timothy Miller, Xiyu Ding et al.

Automatic phenotyping is a task of identifying cohorts of patients that match a predefined set of criteria. Phenotyping typically involves classifying long clinical documents that contain thousands of tokens. At the same time, recent state-of-art transformer-based pre-trained language models limit the input to a few hundred tokens (e.g. 512 tokens for BERT). We evaluate several strategies for incorporating pre-trained sentence encoders into document-level representations of clinical text, and find that hierarchical transformers without pre-training are competitive with task pre-trained models.

Chengyuan Li, Jun Liu, Hailong Hong et al.

Taking the deep learning-based algorithms into account has become a crucial way to boost object detection performance in aerial images. While various neural network representations have been developed, previous works are still inefficient to investigate the noise-resilient performance, especially on aerial images with noise taken by the cameras with telephoto lenses, and most of the research is concentrated in the field of denoising. Of course, denoising usually requires an additional computational burden to obtain higher quality images, while noise-resilient is more of a description of the robustness of the network itself to different noises, which is an attribute of the algorithm itself. For this reason, the work will be started by analyzing the noise-resilient performance of the neural network, and then propose two hypotheses to build a noise-resilient structure. Based on these hypotheses, we compare the noise-resilient ability of the Oct-ResNet with frequency division processing and the commonly used ResNet. In addition, previous feature pyramid networks used for aerial object detection tasks are not specifically designed for the frequency division feature maps of the Oct-ResNet, and they usually lack attention to bridging the semantic gap between diverse feature maps from different depths. On the basis of this, a novel octave convolution-based semantic attention feature pyramid network (OcSaFPN) is proposed to get higher accuracy in object detection with noise. The proposed algorithm tested on three datasets demonstrates that the proposed OcSaFPN achieves a state-of-the-art detection performance with Gaussian noise or multiplicative noise. In addition, more experiments have proved that the OcSaFPN structure can be easily added to existing algorithms, and the noise-resilient ability can be effectively improved.

Chenjie Wang, Bin Luo, Yun Zhang et al.

Most SLAM algorithms are based on the assumption that the scene is static. However, in practice, most scenes are dynamic which usually contains moving objects, these methods are not suitable. In this paper, we introduce DymSLAM, a dynamic stereo visual SLAM system being capable of reconstructing a 4D (3D + time) dynamic scene with rigid moving objects. The only input of DymSLAM is stereo video, and its output includes a dense map of the static environment, 3D model of the moving objects and the trajectories of the camera and the moving objects. We at first detect and match the interesting points between successive frames by using traditional SLAM methods. Then the interesting points belonging to different motion models (including ego-motion and motion models of rigid moving objects) are segmented by a multi-model fitting approach. Based on the interesting points belonging to the ego-motion, we are able to estimate the trajectory of the camera and reconstruct the static background. The interesting points belonging to the motion models of rigid moving objects are then used to estimate their relative motion models to the camera and reconstruct the 3D models of the objects. We then transform the relative motion to the trajectories of the moving objects in the global reference frame. Finally, we then fuse the 3D models of the moving objects into the 3D map of the environment by considering their motion trajectories to obtain a 4D (3D+time) sequence. DymSLAM obtains information about the dynamic objects instead of ignoring them and is suitable for unknown rigid objects. Hence, the proposed system allows the robot to be employed for high-level tasks, such as obstacle avoidance for dynamic objects. We conducted experiments in a real-world environment where both the camera and the objects were moving in a wide range.

Xin Su, Shangqi Guo, Feng Chen

The construction of artificial general intelligence (AGI) was a long-term goal of AI research aiming to deal with the complex data in the real world and make reasonable judgments in various cases like a human. However, the current AI creations, referred to as "Narrow AI", are limited to a specific problem. The constraints come from two basic assumptions of data, which are independent and identical distributed samples and single-valued mapping between inputs and outputs. We completely break these constraints and develop the subjectivity learning theory for general intelligence. We assign the mathematical meaning for the philosophical concept of subjectivity and build the data representation of general intelligence. Under the subjectivity representation, then the global risk is constructed as the new learning goal. We prove that subjectivity learning holds a lower risk bound than traditional machine learning. Moreover, we propose the principle of empirical global risk minimization (EGRM) as the subjectivity learning process in practice, establish the condition of consistency, and present triple variables for controlling the total risk bound. The subjectivity learning is a novel learning theory for unconstrained real data and provides a path to develop AGI.

Mira Rizkallah, Xin Su, Thomas Maugey et al.

The paper addresses the problem of energy compaction of dense 4D light fields by designing geometry-aware local graph-based transforms. Local graphs are constructed on super-rays that can be seen as a grouping of spatially and geometry-dependent angularly correlated pixels. Both non separable and separable transforms are considered. Despite the local support of limited size defined by the super-rays, the Laplacian matrix of the non separable graph remains of high dimension and its diagonalization to compute the transform eigen vectors remains computationally expensive. To solve this problem, we then perform the local spatio-angular transform in a separable manner. We show that when the shape of corresponding super-pixels in the different views is not isometric, the basis functions of the spatial transforms are not coherent, resulting in decreased correlation between spatial transform coefficients. We hence propose a novel transform optimization method that aims at preserving angular correlation even when the shapes of the super-pixels are not isometric. Experimental results show the benefit of the approach in terms of energy compaction. A coding scheme is also described to assess the rate-distortion perfomances of the proposed transforms and is compared to state of the art encoders namely HEVC and JPEG Pleno VM 1.1.