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.

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’s 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 of obtained features. 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. The code for this paper is available at https://github.com/pyashy/SAE_ATD.

Measuring how real images look is a complex task in artificial intelligence research. For example, an image of Albert Einstein holding a smartphone violates common-sense because modern smartphone were invented after Einstein’s death. We introduce a novel method, which we called Through the Looking Glass (TLG), to assess image common sense consistency using Large Vision-Language Models (LVLMs) and Transformer-based encoder. By leveraging LVLM to extract atomic facts from these images, we obtain a mix of accurate facts. We proceed by fine-tuning a compact attention-pooling classifier over encoded atomic facts. Our TLG has achieved a new state-of-the-art performance on the WHOOPS! and WEIRD datasets while leveraging a compact fine-tuning component.

This paper reveals a novel linear characteristic exclusive to transformer decoders, including models like GPT, LLaMA, OPT, BLOOM and others. We analyze embedding transformations between sequential layers, uncovering an almost perfect linear relationship (Procrustes similarity score of 0.99). However, linearity decreases when the residual component is removed, due to a consistently low transformer layer output norm. Our experiments show that pruning or linearly approximating some of the layers does not impact loss or model performance significantly. Moreover, we introduce a cosine-similarity-based regularization in our pretraining experiments on smaller models, aimed at reducing layer linearity. This regularization not only improves performance metrics on benchmarks like Tiny Stories and SuperGLUE but 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 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. This fact 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.

We introduce OmniDialog — the first trimodal comprehensive benchmark grounded in a knowledge graph (Wikidata) to evaluate the generalization of Large Multimodal Models (LMMs) across three modalities. Our benchmark consists of more than 4,000 dialogues, each averaging 10 turns, all annotated and cross-validated by human experts. The dialogues in our dataset are designed to prevent shortcut learning by incorporating various formats and misleading or irrelevant multimodal cues. We also evaluate both multimodal and unimodal models to gain insights into how they process modality inputs introduced in the conversation.

Our research focuses on the most prevalent type of queries— simple questions —exemplified by questions like “What is the capital of France?”. These questions reference an entity such as “France”, which is directly connected (one hop) to the answer entity “Paris” in the underlying knowledge graph (KG). We propose a multilingual Knowledge Graph Question Answering (KGQA) technique that orders potential responses based on the distance between the question’s text embeddings and the answer’s graph embeddings. A system incorporating this novel method is also described in our work. Through comprehensive experimentation using various English and multilingual datasets and two KGs — Freebase and Wikidata — we illustrate the comparative advantage of the proposed method across diverse KG embeddings and languages. This edge is apparent even against robust baseline systems, including seq2seq QA models, search-based solutions and intricate rule-based pipelines. Interestingly, our research underscores that even advanced AI systems like ChatGPT encounter difficulties when tasked with answering simple questions. This finding emphasizes the relevance and effectiveness of our approach, which consistently outperforms such systems. We are making the source code and trained models from our study publicly accessible to promote further advancements in multilingual KGQA.

Text-to-image generation is a significant domain in modern computer vision and achieved substantial improvements through the evolution of generative architectures. Among these, diffusion-based models demonstrated essential quality enhancements. These models generally split into two categories: pixel-level and latent-level approaches. We present Kandinsky – a novel exploration of latent diffusion architecture, combining the principles of image prior models with latent diffusion techniques. The image prior model, is trained separately to map CLIP text and image embeddings. Another distinct feature of the proposed model is the modified MoVQ implementation, which serves as the image autoencoder component. Overall the designed model contains 3.3B parameters. We also deployed a user-friendly demo system that supports diverse generative modes such as text-to-image generation, image fusion, text and image fusion, image variations generation and text-guided inpainting/outpainting. Additionally we released the source code and checkpoints for Kandinsky models. Experimental evaluations demonstrate FID score of 8.03 on the COCO-30K dataset, marking our model as the top open source performer in terms of measurable image generation quality.

Knowledge Graphs (KGs) are symbolically structured storages of facts. The KG embedding contains concise data used in NLP tasks requiring implicit information about the real world. Furthermore, the size of KGs that may be useful in actual NLP assignments is enormous, and creating embedding over it has memory cost issues. We represent KG as a 3rd-order binary tensor and move beyond the standard CP decomposition (CITATION) by using a data-specific generalized version of it (CITATION). The generalization of the standard CP-ALS algorithm allows obtaining optimization gradients without a backpropagation mechanism. It reduces the memory needed in training while providing computational benefits. We propose a MEKER, a memory-efficient KG embedding model, which yields SOTA-comparable performance on link prediction tasks and KG-based Question Answering.

Pixel-level autoregression with Transformer models (Image GPT or iGPT) is one of the recent approaches to image generation that has not received massive attention and elaboration due to quadratic complexity of attention as it imposes huge memory requirements and thus restricts the resolution of the generated images. In this paper, we propose to tackle this problem by adopting Byte-Pair-Encoding (BPE) originally proposed for text processing to the image domain to drastically reduce the length of the modeled sequence. The obtained results demonstrate that it is possible to decrease the amount of computation required to generate images pixel-by-pixel while preserving their quality and the expressiveness of the features extracted from the model. Our results show that there is room for improvement for iGPT-like models with more thorough research on the way to the optimal sequence encoding techniques for images.

In this paper, we present a system for the solution of the cross-lingual and multilingual word-in-context disambiguation task. Task organizers provided monolingual data in several languages, but no cross-lingual training data were available. To address the lack of the officially provided cross-lingual training data, we decided to generate such data ourselves. We describe a simple yet effective approach based on machine translation and back translation of the lexical units to the original language used in the context of this shared task. In our experiments, we used a neural system based on the XLM-R, a pre-trained transformer-based masked language model, as a baseline. We show the effectiveness of the proposed approach as it allows to substantially improve the performance of this strong neural baseline model. In addition, in this study, we present multiple types of the XLM-R based classifier, experimenting with various ways of mixing information from the first and second occurrences of the target word in two samples.