Zijun Wu

Jiarui Zhang, Yuliang Liu, Zijun Wu et al.

Document parsing is a core task in document intelligence, supporting applications such as information extraction, retrieval-augmented generation, and automated document analysis. However, real-world documents often feature complex layouts with multi-level tables, embedded images or formulas, and cross-page structures, which remain challenging for existing OCR systems. We introduce MonkeyOCR v1.5, a unified vision-language framework that enhances both layout understanding and content recognition through a two-stage pipeline. The first stage employs a large multimodal model to jointly predict layout and reading order, leveraging visual information to ensure sequential consistency. The second stage performs localized recognition of text, formulas, and tables within detected regions, maintaining high visual fidelity while reducing error propagation. To address complex table structures, we propose a visual consistency-based reinforcement learning scheme that evaluates recognition quality via render-and-compare alignment, improving structural accuracy without manual annotations. Additionally, two specialized modules, Image-Decoupled Table Parsing and Type-Guided Table Merging, are introduced to enable reliable parsing of tables containing embedded images and reconstruction of tables crossing pages or columns. Comprehensive experiments on OmniDocBench v1.5 demonstrate that MonkeyOCR v1.5 achieves state-of-the-art performance, outperforming PPOCR-VL and MinerU 2.5 while showing exceptional robustness in visually complex document scenarios. A trial link can be found at https://github.com/Yuliang-Liu/MonkeyOCR .

Yangyi Liu, Huan Liu, Liangyan Li et al.

Recent years have witnessed an increased interest in image dehazing. Many deep learning methods have been proposed to tackle this challenge, and have made significant accomplishments dealing with homogeneous haze. However, these solutions cannot maintain comparable performance when they are applied to images with non-homogeneous haze, e.g., NH-HAZE23 dataset introduced by NTIRE challenges. One of the reasons for such failures is that non-homogeneous haze does not obey one of the assumptions that is required for modeling homogeneous haze. In addition, a large number of pairs of non-homogeneous hazy image and the clean counterpart is required using traditional end-to-end training approaches, while NH-HAZE23 dataset is of limited quantities. Although it is possible to augment the NH-HAZE23 dataset by leveraging other non-homogeneous dehazing datasets, we observe that it is necessary to design a proper data-preprocessing approach that reduces the distribution gaps between the target dataset and the augmented one. This finding indeed aligns with the essence of data-centric AI. With a novel network architecture and a principled data-preprocessing approach that systematically enhances data quality, we present an innovative dehazing method. Specifically, we apply RGB-channel-wise transformations on the augmented datasets, and incorporate the state-of-the-art transformers as the backbone in the two-branch framework. We conduct extensive experiments and ablation study to demonstrate the effectiveness of our proposed method.

Zijun Wu, Anup Anand Deshmukh, Yongkang Wu et al.

In Natural Language Processing (NLP), predicting linguistic structures, such as parsing and chunking, has mostly relied on manual annotations of syntactic structures. This paper introduces an unsupervised approach to chunking, a syntactic task that involves grouping words in a non-hierarchical manner. We present a Hierarchical Recurrent Neural Network (HRNN) designed to model word-to-chunk and chunk-to-sentence compositions. Our approach involves a two-stage training process: pretraining with an unsupervised parser and finetuning on downstream NLP tasks. Experiments on multiple datasets reveal a notable improvement of unsupervised chunking performance in both pretraining and finetuning stages. Interestingly, we observe that the emergence of the chunking structure is transient during the neural model's downstream-task training. This study contributes to the advancement of unsupervised syntactic structure discovery and opens avenues for further research in linguistic theory.

Zijun Wu, Yongkang Wu, Lili Mou

Prompt tuning in natural language processing (NLP) has become an increasingly popular method for adapting large language models to specific tasks. However, the transferability of these prompts, especially continuous prompts, between different models remains a challenge. In this work, we propose a zero-shot continuous prompt transfer method, where source prompts are encoded into relative space and the corresponding target prompts are searched for transferring to target models. Experimental results confirm the effectiveness of our method, showing that 'task semantics' in continuous prompts can be generalized across various language models. Moreover, we find that combining 'task semantics' from multiple source models can further enhance the generalizability of transfer.

Yuxing Chen, Guoqing Luo, Zijun Wu et al.

Large Language Models (LLMs) exhibit significant social biases that can perpetuate harmful stereotypes and unfair outcomes. In this paper, we propose Multi-Persona Thinking (MPT), a novel inference-time framework that leverages dialectical reasoning from multiple perspectives to reduce bias. MPT guides models to adopt contrasting social identities (e.g., male and female) along with a neutral viewpoint, and then engages these personas iteratively to expose and correct biases. Through a dialectical reasoning process, the framework transforms the potential weakness of persona assignment into a strength for bias mitigation. We evaluate MPT on two widely used bias benchmarks across both open-source and closed-source models of varying scales. Our results demonstrate substantial improvements over existing prompting-based strategies: MPT achieves the lowest bias while maintaining core reasoning ability.

Xiang Zhang, Senyu Li, Ning Shi et al.

Recent developments in multimodal methodologies have marked the beginning of an exciting era for models adept at processing diverse data types, encompassing text, audio, and visual content. Models like GPT-4V, which merge computer vision with advanced language processing, exhibit extraordinary proficiency in handling intricate tasks that require a simultaneous understanding of both textual and visual information. Prior research efforts have meticulously evaluated the efficacy of these Vision Large Language Models (VLLMs) in various domains, including object detection, image captioning, and other related fields. However, existing analyses have often suffered from limitations, primarily centering on the isolated evaluation of each modality's performance while neglecting to explore their intricate cross-modal interactions. Specifically, the question of whether these models achieve the same level of accuracy when confronted with identical task instances across different modalities remains unanswered. In this study, we take the initiative to delve into the interaction and comparison among these modalities of interest by introducing a novel concept termed cross-modal consistency. Furthermore, we propose a quantitative evaluation framework founded on this concept. Our experimental findings, drawn from a curated collection of parallel vision-language datasets developed by us, unveil a pronounced inconsistency between the vision and language modalities within GPT-4V, despite its portrayal as a unified multimodal model. Our research yields insights into the appropriate utilization of such models and hints at potential avenues for enhancing their design.

Zijun Wu, Bingyuan Liu, Ran Yan et al.

Recent advancements in Large Language Models (LLMs) have significantly enhanced their capacity to process long contexts. However, effectively utilizing this long context remains a challenge due to the issue of distraction, where irrelevant information dominates lengthy contexts, causing LLMs to lose focus on the most relevant segments. To address this, we propose a novel training method that enhances LLMs' ability to discern relevant information through a unique combination of retrieval-based data augmentation and contrastive learning. Specifically, during fine-tuning with long contexts, we employ a retriever to extract the most relevant segments, serving as augmented inputs. We then introduce an auxiliary contrastive learning objective to explicitly ensure that outputs from the original context and the retrieved sub-context are closely aligned. Extensive experiments on long single-document and multi-document QA benchmarks demonstrate the effectiveness of our proposed method.

Zijun Wu, Yongchang Hao, Lili Mou

Large language models achieve state-of-the-art performance but are costly to fine-tune due to their size. Parameter-efficient fine-tuning methods, such as prompt tuning, address this by reducing trainable parameters while maintaining strong performance. However, prior methods tie prompt embeddings to the model's dimensionality, which may not scale well with larger LLMs and more customized LLMs. In this paper, we propose Ultra-Low-dimensional Prompt Tuning (ULPT), which optimizes prompts in a low-dimensional space (e.g., 2D) and use a random but frozen matrix for the up-projection. To enhance alignment, we introduce learnable shift and scale embeddings. ULPT drastically reduces the trainable parameters, e.g., 2D only using 2% parameters compared with vanilla prompt tuning while retaining most of the performance across 21 NLP tasks. Our theoretical analysis shows that random projections can capture high-rank structures effectively, and experimental results demonstrate ULPT's competitive performance over existing parameter-efficient methods.

Hanming Wang, Yunlong Li, Zijun Wu et al.

In the past few years, channel-wise and spatial-wise attention blocks have been widely adopted as supplementary modules in deep neural networks, enhancing network representational abilities while introducing low complexity. Most attention modules follow a squeeze-and-excitation paradigm. However, to design such attention modules, requires a substantial amount of experiments and computational resources. Neural Architecture Search (NAS), meanwhile, is able to automate the design of neural networks and spares the numerous experiments required for an optimal architecture. This motivates us to design a search architecture that can automatically find near-optimal attention modules through NAS. We propose SASE, a Searching Architecture for Squeeze and Excitation operations, to form a plug-and-play attention block by searching within certain search space. The search space is separated into 4 different sets, each corresponds to the squeeze or excitation operation along the channel or spatial dimension. Additionally, the search sets include not only existing attention blocks but also other operations that have not been utilized in attention mechanisms before. To the best of our knowledge, SASE is the first attempt to subdivide the attention search space and search for architectures beyond currently known attention modules. The searched attention module is tested with extensive experiments across a range of visual tasks. Experimental results indicate that visual backbone networks (ResNet-50/101) using the SASE attention module achieved the best performance compared to those using the current state-of-the-art attention modules. Codes are included in the supplementary material, and they will be made public later.

Zijun Wu, Yongchang Hao, Lili Mou

Large language models rely heavily on prompts to specify tasks, recall knowledge and guide reasoning. However, this reliance is inefficient as prompts must be re-read at each step, scale poorly across tasks, and lack mechanisms for modular reuse. We introduce TokMem, a tokenized procedural memory that stores recurring procedures as compact, trainable embeddings. Each memory token encodes both an address to a procedure and a control signal that steers generation, enabling targeted behavior with constant-size overhead. To support continual adaptation, TokMem keeps the backbone model frozen, allowing new procedures to be added without interfering with existing ones. We evaluate TokMem on 1,000 tasks for atomic recall, and on function-calling tasks for compositional recall, where it consistently outperforms retrieval-augmented generation while avoiding repeated context overhead, and fine-tuning with far fewer parameters. These results establish TokMem as a scalable and modular alternative to prompt engineering and fine-tuning, offering an explicit procedural memory for LLMs.

Ning Shi, Zijun Wu

Recent studies have demonstrated the potential to control paraphrase generation, such as through syntax, which has broad applications in various downstream tasks. However, these methods often require detailed parse trees or syntactic exemplars, countering human-like paraphrasing behavior in language use. Furthermore, an inference gap exists, as control specifications are only available during training but not during inference. In this work, we propose a new setup for controlled paraphrase generation. Specifically, we represent user intent as action tokens, embedding and concatenating them with text embeddings, thus flowing together into a self-attention encoder for representation fusion. To address the inference gap, we introduce an optional action token as a placeholder that encourages the model to determine the appropriate action independently when users' intended actions are not provided. Experimental results show that our method successfully enables precise action-controlled paraphrasing and preserves or even enhances performance compared to conventional uncontrolled methods when actions are not given. Our findings promote the concept of action-controlled paraphrasing for a more user-centered design.

Xiang Zhang, Senyu Li, Zijun Wu et al.

Recent advancements in multimodal techniques open exciting possibilities for models excelling in diverse tasks involving text, audio, and image processing. Models like GPT-4V, blending computer vision and language modeling, excel in complex text and image tasks. Numerous prior research endeavors have diligently examined the performance of these Vision Large Language Models (VLLMs) across tasks like object detection, image captioning and others. However, these analyses often focus on evaluating the performance of each modality in isolation, lacking insights into their cross-modal interactions. Specifically, questions concerning whether these vision-language models execute vision and language tasks consistently or independently have remained unanswered. In this study, we draw inspiration from recent investigations into multilingualism and conduct a comprehensive analysis of model's cross-modal interactions. We introduce a systematic framework that quantifies the capability disparities between different modalities in the multi-modal setting and provide a set of datasets designed for these evaluations. Our findings reveal that models like GPT-4V tend to perform consistently modalities when the tasks are relatively simple. However, the trustworthiness of results derived from the vision modality diminishes as the tasks become more challenging. Expanding on our findings, we introduce "Vision Description Prompting," a method that effectively improves performance in challenging vision-related tasks.

Zhize Wu, Huanyi Li, Xiaofeng Wang et al.

Household garbage images are usually faced with complex backgrounds, variable illuminations, diverse angles, and changeable shapes, which bring a great difficulty in garbage image classification. Due to the ability to discover problem-specific features, deep learning and especially convolutional neural networks (CNNs) have been successfully and widely used for image representation learning. However, available and stable household garbage datasets are insufficient, which seriously limits the development of research and application. Besides, the state of the art in the field of garbage image classification is not entirely clear. To solve this problem, in this study, we built a new open benchmark dataset for household garbage image classification by simulating different lightings, backgrounds, angles, and shapes. This dataset is named 30 Classes of Household Garbage Images (HGI-30), which contains 18,000 images of 30 household garbage classes. The publicly available HGI-30 dataset allows researchers to develop accurate and robust methods for household garbage recognition. We also conducted experiments and performance analysis of the state-of-the-art deep CNN methods on HGI-30, which serves as baseline results on this benchmark.

Zijun Wu, Zi Xuan Zhang, Atharva Naik et al.

Natural language inference (NLI) aims to determine the logical relationship between two sentences, such as Entailment, Contradiction, and Neutral. In recent years, deep learning models have become a prevailing approach to NLI, but they lack interpretability and explainability. In this work, we address the explainability of NLI by weakly supervised logical reasoning, and propose an Explainable Phrasal Reasoning (EPR) approach. Our model first detects phrases as the semantic unit and aligns corresponding phrases in the two sentences. Then, the model predicts the NLI label for the aligned phrases, and induces the sentence label by fuzzy logic formulas. Our EPR is almost everywhere differentiable and thus the system can be trained end to end. In this way, we are able to provide explicit explanations of phrasal logical relationships in a weakly supervised manner. We further show that such reasoning results help textual explanation generation.

Xiaohong Liu, Zhihao Shi, Zijun Wu et al.

We propose an enhanced multi-scale network, dubbed GridDehazeNet+, for single image dehazing. The proposed dehazing method does not rely on the Atmosphere Scattering Model (ASM), and an explanation as to why it is not necessarily performing the dimension reduction offered by this model is provided. GridDehazeNet+ consists of three modules: pre-processing, backbone, and post-processing. The trainable pre-processing module can generate learned inputs with better diversity and more pertinent features as compared to those derived inputs produced by hand-selected pre-processing methods. The backbone module implements multi-scale estimation with two major enhancements: 1) a novel grid structure that effectively alleviates the bottleneck issue via dense connections across different scales; 2) a spatial-channel attention block that can facilitate adaptive fusion by consolidating dehazing-relevant features. The post-processing module helps to reduce the artifacts in the final output. Due to domain shift, the model trained on synthetic data may not generalize well on real data. To address this issue, we shape the distribution of synthetic data to match that of real data, and use the resulting translated data to finetune our network. We also propose a novel intra-task knowledge transfer mechanism that can memorize and take advantage of synthetic domain knowledge to assist the learning process on the translated data. Experimental results demonstrate that the proposed method outperforms the state-of-the-art on several synthetic dehazing datasets, and achieves the superior performance on real-world hazy images after finetuning.

Thomas Weise, Zijun Wu, Markus Wagner

A commonly used strategy for improving optimization algorithms is to restart the algorithm when it is believed to be trapped in an inferior part of the search space. Building on the recent success of Bet-and-Run approaches for restarted local search solvers, we introduce an improved generic Bet-and-Run strategy. The goal is to obtain the best possible results within a given time budget t using a given black-box optimization algorithm. If no prior knowledge about problem features and algorithm behavior is available, the question about how to use the time budget most efficiently arises. We propose to first start k>=1 independent runs of the algorithm during an initialization budget t1<t, pausing these runs, then apply a decision maker D to choose 1<=m<=k runs from them (consuming t2>=0 time units in doing so), and then continuing these runs for the remaining t3=t-t1-t2 time units. In previous Bet-and-Run strategies, the decision maker D=currentBest would simply select the run with the best- so-far results at negligible time. We propose using more advanced methods to discriminate between "good" and "bad" sample runs, with the goal of increasing the correlation of the chosen run with the a-posteriori best one. We test several different approaches, including neural networks trained or polynomials fitted on the current trace of the algorithm to predict which run may yield the best results if granted the remaining budget. We show with extensive experiments that this approach can yield better results than the previous methods, but also find that the currentBest method is a very reliable and robust baseline approach.

Zijun Wu, Rolf Moehring, Jianhui Lai

This article analyzes the stochastic runtime of a Cross-Entropy Algorithm on two classes of traveling salesman problems. The algorithm shares main features of the famous Max-Min Ant System with iteration-best reinforcement. For simple instances that have a $\{1,n\}$-valued distance function and a unique optimal solution, we prove a stochastic runtime of $O(n^{6+ε})$ with the vertex-based random solution generation, and a stochastic runtime of $O(n^{3+ε}\ln n)$ with the edge-based random solution generation for an arbitrary $ε\in (0,1)$. These runtimes are very close to the known expected runtime for variants of Max-Min Ant System with best-so-far reinforcement. They are obtained for the stronger notion of stochastic runtime, which means that an optimal solution is obtained in that time with an overwhelming probability, i.e., a probability tending exponentially fast to one with growing problem size. We also inspect more complex instances with $n$ vertices positioned on an $m\times m$ grid. When the $n$ vertices span a convex polygon, we obtain a stochastic runtime of $O(n^{3}m^{5+ε})$ with the vertex-based random solution generation, and a stochastic runtime of $O(n^{2}m^{5+ε})$ for the edge-based random solution generation. When there are $k = O(1)$ many vertices inside a convex polygon spanned by the other $n-k$ vertices, we obtain a stochastic runtime of $O(n^{4}m^{5+ε}+n^{6k-1}m^ε)$ with the vertex-based random solution generation, and a stochastic runtime of $O(n^{3}m^{5+ε}+n^{3k}m^ε)$ with the edge-based random solution generation. These runtimes are better than the expected runtime for the so-called $(μ\!+\!λ)$ EA reported in a recent article, and again obtained for the stronger notion of stochastic runtime.