Artificial Text Detection (ATD) is becoming increasingly important with the rise of advanced Large Language Models (LLMs). Despite numerous efforts, no single algorithm performs consistently well across different types of unseen text or guarantees effective generalization to new LLMs. Interpretability plays a crucial role in achieving this goal. In this study, we enhance ATD interpretability by using Sparse Autoencoders (SAE) to extract features from Gemma-2-2b residual stream. We identify both interpretable and efficient features, analyzing their semantics and relevance through domain- and model-specific statistics, a steering approach, and manual or LLM-based interpretation. Our methods offer valuable insights into how texts from various models differ from human-written content. We show that modern LLMs have a distinct writing style, especially in information-dense domains, even though they can produce human-like outputs with personalized prompts.

We introduce methods to quantify how Large Language Models (LLMs) encode and store contextual information, revealing that tokens often seen as minor (e.g., determiners, punctuation) carry surprisingly high context. Notably, removing these tokens -- especially stopwords, articles, and commas -- consistently degrades performance on MMLU and BABILong-4k, even if removing only irrelevant tokens. Our analysis also shows a strong correlation between contextualization and linearity, where linearity measures how closely the transformation from one layer's embeddings to the next can be approximated by a single linear mapping. These findings underscore the hidden importance of filler tokens in maintaining context. For further exploration, we present LLM-Microscope, an open-source toolkit that assesses token-level nonlinearity, evaluates contextual memory, visualizes intermediate layer contributions (via an adapted Logit Lens), and measures the intrinsic dimensionality of representations. This toolkit illuminates how seemingly trivial tokens can be critical for long-range understanding.

We propose a universal adversarial attack on multimodal Large Language Models (LLMs) that leverages a single optimized image to override alignment safeguards across diverse queries and even multiple models. By backpropagating through the vision encoder and language head, we craft a synthetic image that forces the model to respond with a targeted phrase (e.g., ''Sure, here it is'') or otherwise unsafe content-even for harmful prompts. In experiments on the SafeBench benchmark, our method achieves significantly higher attack success rates than existing baselines, including text-only universal prompts (e.g., up to 93% on certain models). We further demonstrate cross-model transferability by training on several multimodal LLMs simultaneously and testing on unseen architectures. Additionally, a multi-answer variant of our approach produces more natural-sounding (yet still malicious) responses. These findings underscore critical vulnerabilities in current multimodal alignment and call for more robust adversarial defenses. We will release code and datasets under the Apache-2.0 license. Warning: some content generated by Multimodal LLMs in this paper may be offensive to some readers.

In this paper we present an approach to reduce hallucinations in Large Language Models (LLMs) by incorporating Knowledge Graphs (KGs) as an additional modality. Our method involves transforming input text into a set of KG embeddings and using an adapter to integrate these embeddings into the language model space, without relying on external retrieval processes. To facilitate this, we created WikiEntities, a dataset containing over 3 million Wikipedia texts annotated with entities from Wikidata and their corresponding embeddings from PyTorch-BigGraph. This dataset serves as a valuable resource for training Entity Linking models and adapting the described method to various LLMs using specialized adapters. Our method does not require fine-tuning of the language models themselves; instead, we only train the adapter. This ensures that the model's performance on other tasks is not affected. We trained an adapter for the Mistral 7B, LLaMA 2-7B (chat), and LLaMA 3-8B (instruct) models using this dataset and demonstrated that our approach improves performance on the HaluEval, True-False benchmarks and FEVER dataset. The results indicate that incorporating KGs as a new modality can effectively reduce hallucinations and improve the factual accuracy of language models, all without the need for external retrieval.

This paper reveals a novel linear characteristic exclusive to transformer decoders, including models such as GPT, LLaMA, OPT, BLOOM and others. We analyze embedding transformations between sequential layers, uncovering a near-perfect linear relationship (Procrustes similarity score of 0.99). However, linearity decreases when the residual component is removed due to a consistently low output norm of the transformer layer. Our experiments show that removing or linearly approximating some of the most linear blocks of transformers does not affect significantly the loss or model performance. Moreover, in our pretraining experiments on smaller models we introduce a cosine-similarity-based regularization, aimed at reducing layer linearity. This regularization improves performance metrics on benchmarks like Tiny Stories and SuperGLUE and as well successfully decreases the linearity of the models. This study challenges the existing understanding of transformer architectures, suggesting that their operation may be more linear than previously assumed.

In recent years, multimodal architectures emerged as a powerful paradigm for enhancing artificial intelligence (AI) systems, enabling them to process and understand multiple types of data simultaneously [1, 2, 3]. The integration of different data modalities, such as text and images, has significantly improved the capabilities of large language models (LLMs) in various tasks, ranging from visual question answering (VQA) [4] to complex decision-making processes [5, 6]. However, the challenge of effectively coupling various data types remains a significant obstacle in the development of truly integrative AI models. Furthermore, such multimodal multitask architectures are interpreted as the first steps towards the development of the artificial general intelligence (AGI), expanding the number of challenges in world cognition. This work introduces the OmniFusion model, a novel multimodal architecture that leverages the strengths of pretrained LLMs and introduces specialized adapters for processing visual information.

In this study, we present an investigation into the anisotropy dynamics and intrinsic dimension of embeddings in transformer architectures, focusing on the dichotomy between encoders and decoders. Our findings reveal that the anisotropy profile in transformer decoders exhibits a distinct bell-shaped curve, with the highest anisotropy concentrations in the middle layers. This pattern diverges from the more uniformly distributed anisotropy observed in encoders. In addition, we found that the intrinsic dimension of embeddings increases in the initial phases of training, indicating an expansion into higher-dimensional space. Which is then followed by a compression phase towards the end of training with dimensionality decrease, suggesting a refinement into more compact representations. Our results provide fresh insights to the understanding of encoders and decoders embedding properties.

Pre-trained Text-to-Text Language Models (LMs), such as T5 or BART yield promising results in the Knowledge Graph Question Answering (KGQA) task. However, the capacity of the models is limited and the quality decreases for questions with less popular entities. In this paper, we present a novel approach which works on top of the pre-trained Text-to-Text QA system to address this issue. Our simple yet effective method performs filtering and re-ranking of generated candidates based on their types derived from Wikidata "instance_of" property.