Real-world document question-answering that relies on screenshots, such as bills and forms, requires evidence that is often spatially localised and visually cluttered. However, most Screenshot Language Models (SLMs) encode the entire page holistically and rely on implicit attention to ''find'' relevant content, which limits both accuracy and efficiency. We present RegionSLM, a region-aware SLM designed to explicitly connect the question to its supporting regions. RegionSLM has two key components: (1) a patch-relevance router that learns a query–region relevance distribution, enabling the model to produce a box-free relevance prior at inference; and (2) Relevance-Guided Region Pooling (RGRP), a query-conditional attention–pooling module that aggregates dense features into a small set of region tokens, which preserves grounding signals while reducing computational overhead. To support training and evaluation, we further curate ReDoc, a region-supervised corpus with 105k documents and 350k question-answer pairs, obtained via a question-guided two-step filtering procedure. Extensive experiments on 12 datasets demonstrate that explicitly learning query–region relevance and pooling it into compact region tokens is an effective and practical recipe for document retrieval and understanding.

Question Answering (QA) and Information Retrieval (IR) play a crucial role in information-seeking pipelines implemented in many emerging AI research assistant applications. Large Language Models (LLMs) have demonstrated exceptional effectiveness on QA tasks, with Retrieval Augmented Generation (RAG) techniques often boosting the results. However, in many of those emerging applications, the onus of conducting the actual literature search falls on the user, i.e. the user searches for the relevant literature and the LLM-based assistant extracts the solicited answers from each of the user-supplied documents. The interplay between the quality of the user-conducted search and the quality of the final results remains understudied.

Diffusion-based Text-to-Image (T2I) models have demonstrated significant potential in image restoration. However, existing models continue to grapple with challenges such as complex training and prompt design. We introduce a new perspective for improving image restoration by injecting knowledge from pretrained vision-language models into current T2I models. We empirically show that the degradation and content representations in BLIP-2 can be linearly separated, providing promising degradation guidance for image restoration. Specifically, the Feature Difference Instruction (FDI) is first extracted by Q-Formers through a simple subtraction operation based on reference image pairs. Then, we propose a multi-scale FDI adapter to decouple the degradation style and corrupted artifacts, and inject the styleflow exclusively into specific blocks through adapter-tuning, thereby preventing noise interference and eschewing the need for cumbersome weight retraining. In this way, we can train various task-specific adapters according to different degradations, achieving rich detail enhancement in the restoration results. Furthermore, the proposed FDI adapters have attractive properties of practical value, such as composability and generalization ability for all-in-one and mixed-degradation restoration. Extensive experiments under various settings demonstrate that our method has promising repairing quality over 10 image restoration tasks and a wide range of other applications.

In this paper, we address contextual limitations of current deep learning-based and heuristic key phrase extraction tools as applied to the domain of cybersecurity. To address these limitations, we develop a hybrid system that augments state-of-the-art (SOTA) transformers for the task of key phrase sequence labeling, using a novel set of part-of-speech (POS) and role-aware tagging rules to generate fine-grained tag sequences from short text corpora. Next, we fine-tune multiple SOTA deep learning (DL) language model (LM) architectures to these transformed sequences. We then evaluate the architectures by measuring the outcomes from respective LMs to select the best-performing underlying transformers for extracting cybersecurity key phrases. This new ensemble achieves very significant predictive gains over SOTA baselines on general cybersecurity corpora, such as F1 scores at least 25% higher than hybrid SOTA transformers fine-tuned using baseline tagging rules on the generic corpus, with a much less significant tradeoff (of less than 5% in F1) on a vulnerability-specific corpus.

We present an end-to-end transformer-based framework named TRDLU for the task of Document Layout Understanding (DLU). DLU is the fundamental task to automatically understand document structures. To accurately detect content boxes and classify them into semantically meaningful classes from various formats of documents is still an open challenge. Recently, transformer-based detection neural networks have shown their capability over traditional convolutional-based methods in the object detection area. In this paper, we consider DLU as a detection task, and introduce TRDLU which integrates transformer-based vision backbone and transformer encoder-decoder as detection pipeline. TRDLU is only a visual feature-based framework, but its performance is even better than multi-modal feature-based models. To the best of our knowledge, this is the first study of employing a fully transformer-based framework in DLU tasks. We evaluated TRDLU on three different DLU benchmark datasets, each with strong baselines. TRDLU outperforms the current state-of-the-art methods on all of them.

This paper presents a novel use case of Graph Convolutional Network (GCN) learning representations for predictive data mining, specifically from user/task data in the domain of high-performance computing (HPC). It outlines an approach based on a coalesced data set: logs from the Slurm workload manager, joined with user experience survey data from computational cluster users. We introduce a new method of constructing a heterogeneous unweighted HPC graph consisting of multiple typed nodes after revealing the manifold relations between the nodes. The GCN structure used here supports two tasks: i) determining whether a job will complete or fail and ii) predicting memory and CPU requirements by training the GCN semi-supervised classification model and regression models on the generated graph. The graph is partitioned into partitions using graph clustering. We conducted classification and regression experiments using the proposed framework on our HPC log dataset and evaluated predictions by our trained models against baselines using test_score, F1-score, precision, recall for classification, and R1 score for regression, showing that our framework achieves significant improvements.

We present a novel computer vision-based deep learning approach for metadata extraction as both a central component of and an ancillary aid to structured information extraction from scientific literature which has various formats. The number of scientific publications is growing rapidly, but existing methods cannot combine the techniques of layout extraction and text recognition efficiently because of the various formats used by scientific literature publishers. In this paper, we introduce an end-to-end trainable neural network for segmenting and labeling the main regions of scientific documents, while simultaneously recognizing text from the detected regions. The proposed framework combines object detection techniques based on Recurrent Convolutional Neural Network (RCNN) for scientific document layout detection with Convolutional Recurrent Neural Network (CRNN) for text recognition. We also contribute a novel data set of main region annotations for scientific literature metadata information extraction to complement the limited availability of high-quality data set. The final outputs of the network are the text content (payload) and the corresponding labels of the major regions. Our results show that our model outperforms state-of-the-field baselines.

We present an approach for adapting convolutional neural networks for object recognition and classification to scientific literature layout detection (SLLD), a shared subtask of several information extraction problems. Scientific publications contain multiple types of information sought by researchers in various disciplines, organized into an abstract, bibliography, and sections documenting related work, experimental methods, and results; however, there is no effective way to extract this information due to their diverse layout. In this paper, we present a novel approach to developing an end-to-end learning framework to segment and classify major regions of a scientific document. We consider scientific document layout analysis as an object detection task over digital images, without any additional text features that need to be added into the network during the training process. Our technical objective is to implement transfer learning via fine-tuning of pre-trained networks and thereby demonstrate that this deep learning architecture is suitable for tasks that lack very large document corpora for training ab initio. As part of the experimental test bed for empirical evaluation of this approach, we created a merged multi-corpus data set for scientific publication layout detection tasks. Our results show good improvement with fine-tuning of a pre-trained base network using this merged data set, compared to the baseline convolutional neural network architecture.