Hallucinations in large language models are a widespread problem, yet the mechanisms behind whether models will hallucinate are poorly understood, limiting our ability to solve this problem. Using sparse autoencoders as an interpretability tool, we discover that a key part of these mechanisms is entity recognition, where the model detects if an entity is one it can recall facts about. Sparse autoencoders uncover meaningful directions in the representation space, these detect whether the model recognizes an entity, e.g. detecting it doesn't know about an athlete or a movie. This suggests that models can have self-knowledge: internal representations about their own capabilities. These directions are causally relevant: capable of steering the model to refuse to answer questions about known entities, or to hallucinate attributes of unknown entities when it would otherwise refuse. We demonstrate that despite the sparse autoencoders being trained on the base model, these directions have a causal effect on the chat model's refusal behavior, suggesting that chat finetuning has repurposed this existing mechanism. Furthermore, we provide an initial exploration into the mechanistic role of these directions in the model, finding that they disrupt the attention of downstream heads that typically move entity attributes to the final token.

As object detection becomes integral to many safety-critical applications, understanding its vulnerabilities is essential. Backdoor attacks, in particular, pose a significant threat by implanting hidden backdoor in a victim model, which adversaries can later exploit to trigger malicious behaviors during inference. However, current backdoor techniques are limited to static scenarios where attackers must define a malicious objective before training, locking the attack into a predetermined action without inference-time adaptability. Given the expressive output space in object detection, including object existence detection, bounding box estimation, and object classification, the feasibility of implanting a backdoor that provides inference-time control with a high degree of freedom remains unexplored. This paper introduces AnywhereDoor, a flexible backdoor attack tailored for object detection. Once implanted, AnywhereDoor enables adversaries to specify different attack types (object vanishing, fabrication, or misclassification) and configurations (untargeted or targeted with specific classes) to dynamically control detection behavior. This flexibility is achieved through three key innovations: (i) objective disentanglement to support a broader range of attack combinations well beyond what existing methods allow; (ii) trigger mosaicking to ensure backdoor activations are robust, even against those object detectors that extract localized regions from the input image for recognition; and (iii) strategic batching to address object-level data imbalances that otherwise hinders a balanced manipulation. Extensive experiments demonstrate that AnywhereDoor provides attackers with a high degree of control, achieving an attack success rate improvement of nearly 80% compared to adaptations of existing methods for such flexible control.

Camera traps offer enormous new opportunities in ecological studies, but current automated image analysis methods often lack the contextual richness needed to support impactful conservation outcomes. Here we present an integrated approach that combines deep learning-based vision and language models to improve ecological reporting using data from camera traps. We introduce a two-stage system: YOLOv10-X to localise and classify species (mammals and birds) within images, and a Phi-3.5-vision-instruct model to read YOLOv10-X binding box labels to identify species, overcoming its limitation with hard to classify objects in images. Additionally, Phi-3.5 detects broader variables, such as vegetation type, and time of day, providing rich ecological and environmental context to YOLO's species detection output. When combined, this output is processed by the model's natural language system to answer complex queries, and retrieval-augmented generation (RAG) is employed to enrich responses with external information, like species weight and IUCN status (information that cannot be obtained through direct visual analysis). This information is used to automatically generate structured reports, providing biodiversity stakeholders with deeper insights into, for example, species abundance, distribution, animal behaviour, and habitat selection. Our approach delivers contextually rich narratives that aid in wildlife management decisions. By providing contextually rich insights, our approach not only reduces manual effort but also supports timely decision-making in conservation, potentially shifting efforts from reactive to proactive management.

Language models perform differently across languages. It has been previously suggested that morphological typology may explain some of this variability (Cotterell et al., 2018). We replicate previous analyses and find additional new evidence for a performance gap between agglutinative and fusional languages, where fusional languages, such as English, tend to have better language modeling performance than morphologically more complex languages like Turkish. We then propose and test three possible causes for this performance gap: morphological alignment of tokenizers, tokenization quality, and disparities in dataset sizes and measurement. To test the morphological alignment hypothesis, we present MorphScore, a tokenizer evaluation metric, and supporting datasets for 22 languages. We find some evidence that tokenization quality explains the performance gap, but none for the role of morphological alignment. Instead we find that the performance gap is most reduced when training datasets are of equivalent size across language types, but only when scaled according to the so-called "byte-premium" -- the different encoding efficiencies of different languages and orthographies. These results suggest that no language is harder or easier for a language model to learn on the basis of its morphological typology. Differences in performance can be attributed to disparities in dataset size. These results bear on ongoing efforts to improve performance for low-performing and under-resourced languages.

Graph neural networks (GNNs) have emerged as powerful models for learning representations of graph data showing state of the art results in various tasks. Nevertheless, the superiority of these methods is usually supported by either evaluating their performance on small subset of benchmark datasets or by reasoning about their expressive power in terms of certain graph isomorphism tests. In this paper we critically analyse both these aspects through a transductive setting for the task of node classification. First, we delve deeper into the case of multi-label node classification which offers a more realistic scenario and has been ignored in most of the related works. Through analysing the training dynamics for GNN methods we highlight the failure of GNNs to learn over multi-label graph datasets even for the case of abundant training data. Second, we show that specifically for transductive node classification, even the most expressive GNN may fail to learn in absence of node attributes and without using explicit label information as input. To overcome this deficit, we propose a straightforward approach, referred to as GNN-MultiFix, that integrates the feature, label, and positional information of a node. GNN-MultiFix demonstrates significant improvement across all the multi-label datasets. We release our code at https://anonymous.4open.science/r/Graph-MultiFix-4121.

In order to solve the problem of insufficient generation quality caused by traditional patent text abstract generation models only originating from patent specifications, the problem of new terminology OOV caused by rapid patent updates, and the problem of information redundancy caused by insufficient consideration of the high professionalism, accuracy, and uniqueness of patent texts, we proposes a patent text abstract generation model (MSEA) based on a master-slave encoder architecture; Firstly, the MSEA model designs a master-slave encoder, which combines the instructions in the patent text with the claims as input, and fully explores the characteristics and details between the two through the master-slave encoder; Then, the model enhances the consideration of new technical terms in the input sequence based on the pointer network, and further enhances the correlation with the input text by re weighing the "remembered" and "for-gotten" parts of the input sequence from the encoder; Finally, an enhanced repetition suppression mechanism for patent text was introduced to ensure accurate and non redundant abstracts generated. On a publicly available patent text dataset, compared to the state-of-the-art model, Improved Multi-Head Attention Mechanism (IMHAM), the MSEA model achieves an improvement of 0.006, 0.005, and 0.005 in Rouge-1, Rouge-2, and Rouge-L scores, respectively. MSEA leverages the characteristics of patent texts to effectively enhance the quality of patent text generation, demonstrating its advancement and effectiveness in the experiments.

Large language models (LLMs) deliver impressive results but face challenges from increasing model sizes and computational costs. Structured pruning reduces model size and speeds up inference but often causes uneven degradation across domains, leading to biased performance. To address this, we propose DRPruning, which incorporates distributionally robust optimization to restore balanced performance across domains, along with further improvements to enhance robustness. Experiments in monolingual and multilingual settings show that our method surpasses similarly sized models in pruning and continued pretraining over perplexity, downstream tasks, and instruction tuning. We further provide analysis demonstrating the robustness of our method towards various domains and distribution shifts. Furthermore, our method automatically determines optimal reference losses and data ratios, suggesting potential for broader applications. Our code is available at https://github.com/hexuandeng/DRPruning.

Predicting future international events from textual information, such as news articles, has tremendous potential for applications in global policy, strategic decision-making, and geopolitics. However, existing datasets available for this task are often limited in quality, hindering the progress of related research. In this paper, we introduce WORLDREP (WORLD Relationship and Event Prediction), a novel dataset designed to address these limitations by leveraging the advanced reasoning capabilities of large-language models (LLMs). Our dataset features high-quality scoring labels generated through advanced prompt modeling and rigorously validated by domain experts in political science. We showcase the quality and utility of WORLDREP for real-world event prediction tasks, demonstrating its effectiveness through extensive experiments and analysis. Furthermore, we publicly release our dataset along with the full automation source code for data collection, labeling, and benchmarking, aiming to support and advance research in text-based event prediction.

Medical imaging has significantly revolutionized medical diagnostics and treatment planning, progressing from early X-ray usage to sophisticated methods like MRIs, CT scans, and ultrasounds. This paper investigates the use of deep learning for medical image captioning, with a particular focus on uterine ultrasound images. These images are vital in obstetrics and gynecology for diagnosing and monitoring various conditions across different age groups. However, their interpretation is often challenging due to their complexity and variability. To address this, a deep learning-based medical image captioning system was developed, integrating Convolutional Neural Networks with a Bidirectional Gated Recurrent Unit network. This hybrid model processes both image and text features to generate descriptive captions for uterine ultrasound images. Our experimental results demonstrate the effectiveness of this approach over baseline methods, with the proposed model achieving superior performance in generating accurate and informative captions, as indicated by higher BLEU and ROUGE scores. By enhancing the interpretation of uterine ultrasound images, our research aims to assist medical professionals in making timely and accurate diagnoses, ultimately contributing to improved patient care.

Training multimodal generative models on large, uncurated datasets can result in users being exposed to harmful, unsafe and controversial or culturally-inappropriate outputs. While model editing has been proposed to remove or filter undesirable concepts in embedding and latent spaces, it can inadvertently damage learned manifolds, distorting concepts in close semantic proximity. We identify limitations in current model editing techniques, showing that even benign, proximal concepts may become misaligned. To address the need for safe content generation, we propose a modular, dynamic solution that leverages safety-context embeddings and a dual reconstruction process using tunable weighted summation in the latent space to generate safer images. Our method preserves global context without compromising the structural integrity of the learned manifolds. We achieve state-of-the-art results on safe image generation benchmarks, while offering controllable variation of model safety. We identify trade-offs between safety and censorship, which presents a necessary perspective in the development of ethical AI models. We will release our code. Keywords: Text-to-Image Models, Generative AI, Safety, Reliability, Model Editing