The generalization of language models (LMs) is undergoing active debates, contrasting their potential for general intelligence with their struggles with basic knowledge composition (e.g., reverse/transition curse). This paper uncovers the phenomenon of linear correlations in LMs during knowledge composition. For explanation, there exists a linear transformation between certain related knowledge that maps the next token prediction logits from one prompt to another, e.g., "X lives in the city of" $\rightarrow$ "X lives in the country of" for every given X. This mirrors the linearity in human knowledge composition, such as Paris $\rightarrow$ France. Our findings indicate that the linear transformation is resilient to large-scale fine-tuning, generalizing updated knowledge when aligned with real-world relationships, but causing hallucinations when it deviates. Empirical results suggest that linear correlation can serve as a potential identifier of LM's generalization. Finally, we show such linear correlations can be learned with a single feedforward network and pre-trained vocabulary representations, indicating LM generalization heavily relies on the latter.

Data selection has emerged as a crucial downstream application of data valuation. While existing data valuation methods have shown promise in selection tasks, the theoretical foundations and full potential of using data values for selection remain largely unexplored. In this work, we first demonstrate that data values applied for selection can be naturally reformulated as a sequential-decision-making problem, where the optimal data value can be derived through dynamic programming. We show this framework unifies and reinterprets existing methods like Data Shapley through the lens of approximate dynamic programming, specifically as myopic reward function approximations to this sequential problem. Furthermore, we analyze how sequential data selection optimality is affected when the ground-truth utility function exhibits monotonic submodularity with curvature. To address the computational challenges in obtaining optimal data values, we propose an efficient approximation scheme using learned bipartite graphs as surrogate utility models, ensuring greedy selection is still optimal when the surrogate utility is correctly specified and learned. Extensive experiments demonstrate the effectiveness of our approach across diverse datasets.

The goal of this paper is threefold: i) to present some practical aspects of using full-text version of Corpus of Historical American English (COHA), the largest diachronic multi-genre corpus of the English language, in the investigation of a linguistic trend of change; ii) to test a widely held assumption that sentence length in written English has been steadily decreasing over the past few centuries; iii) to point to a possible link between the changes in sentence length and changes in the English syntactic usage. The empirical proof of concept for iii) is provided by the decline in the frequency of the non-finite purpose subordinator in order to. Sentence length, genre and the likelihood of occurrence of in order to are shown to be interrelated.

Length-control summarization aims to condense long texts into a short one within a certain length limit. Previous approaches often use autoregressive (AR) models and treat the length requirement as a soft constraint, which may not always be satisfied. In this study, we propose a novel length-control decoding algorithm based on the Directed Acyclic Transformer (DAT). Our approach allows for multiple plausible sequence fragments and predicts a \emph{path} to connect them. In addition, we propose a Sequence Maximum a Posteriori (SeqMAP) decoding algorithm that marginalizes different possible paths and finds the most probable summary satisfying the length budget. Our algorithm is based on beam search, which further facilitates a reranker for performance improvement. Experimental results on the Gigaword and DUC2004 datasets demonstrate our state-of-the-art performance for length-control summarization.

Fuzzy implication functions are a key area of study in fuzzy logic, extending the classical logical conditional to handle truth degrees in the interval $[0,1]$. While existing literature often focuses on a limited number of families, in the last ten years many new families have been introduced, each defined by specific construction methods and having different key properties. This survey aims to provide a comprehensive and structured overview of the diverse families of fuzzy implication functions, emphasizing their motivations, properties, and potential applications. By organizing the information schematically, this document serves as a valuable resource for both theoretical researchers seeking to avoid redundancy and practitioners looking to select appropriate operators for specific applications.

This report analyzes key trends, challenges, risks, and opp ortunities associated with the development of Large Language Models (LLMs) globally. It examines natio nal experiences in developing LLMs and assesses the feasibility of investment in this sector. Addi tionally, the report explores strategies for implementing, regulating, and financing AI projects at the s tate level. International experiences indicate that LLMs significantl y enhance administrative efficiency. In regulatory processes, they streamline the management of le gal documents (Albania, Serbia), facilitate communication between government authorities and citizen s (Netherlands), and support public procurement and legal translations (Albania).

The field of mechanistic interpretability in pre-trained transformer models has demonstrated substantial evidence supporting the ''linear representation hypothesis'', which is the idea that high level concepts are encoded as vectors in the space of activations of a model. Studies also show that model generation behavior can be steered toward a given concept by adding the concept's vector to the corresponding activations. We show how to leverage these properties to build a form of logical implication into models, enabling transparent and interpretable adjustments that induce a chosen generation behavior in response to the presence of any given concept. Our method, Logical Implication Model Steering (LIMS), unlocks new hand engineered reasoning capabilities by integrating neuro-symbolic logic into pre-trained transformer models.

This study investigates continual fine-tuning strategies for deep learning in online longitudinal electroencephalography (EEG) motor imagery (MI) decoding within a causal setting involving a large user group and multiple sessions per participant. We are the first to explore such strategies across a large user group, as longitudinal adaptation is typically studied in the single-subject setting with a single adaptation strategy, which limits the ability to generalize findings. First, we examine the impact of different fine-tuning approaches on decoder performance and stability. Building on this, we integrate online test-time adaptation (OTTA) to adapt the model during deployment, complementing the effects of prior fine-tuning. Our findings demonstrate that fine-tuning that successively builds on prior subject-specific information improves both performance and stability, while OTTA effectively adapts the model to evolving data distributions across consecutive sessions, enabling calibration-free operation. These results offer valuable insights and recommendations for future research in longitudinal online MI decoding and highlight the importance of combining domain adaptation strategies for improving BCI performance in real-world applications. Clinical Relevance: Our investigation enables more stable and efficient long-term motor imagery decoding, which is critical for neurorehabilitation and assistive technologies.

Implicit neural representations (INRs) exhibit growing promise in addressing Earth representation challenges, ranging from emissions monitoring to climate modeling. However, existing methods disproportionately prioritize global average performance, whereas practitioners require fine-grained insights to understand biases and variations in these models. To bridge this gap, we introduce FAIR-Earth: a first-of-its-kind dataset explicitly crafted to examine and challenge inequities in Earth representations. FAIR-Earth comprises various high-resolution Earth signals and uniquely aggregates extensive metadata along stratifications like landmass size and population density to assess the fairness of models. Evaluating state-of-the-art INRs across the various modalities of FAIR-Earth, we uncover striking performance disparities. Certain subgroups, especially those associated with high-frequency signals (e.g., islands, coastlines), are consistently poorly modeled by existing methods. In response, we propose spherical wavelet encodings, building on previous spatial encoding research. Leveraging the multi-resolution capabilities of wavelets, our encodings yield consistent performance over various scales and locations, offering more accurate and robust representations of the biased subgroups. These open-source contributions represent a crucial step towards the equitable assessment and deployment of Earth INRs.

Recent advancements in AI have revolutionized property prediction in materials science and accelerating material discovery. Graph neural networks (GNNs) stand out due to their ability to represent crystal structures as graphs, effectively capturing local interactions and delivering superior predictions. However, these methods often lose critical global information, such as crystal systems and repetitive unit connectivity. To address this, we propose CAST, a cross-attention-based multimodal fusion model that integrates graph and text modalities to preserve essential material information. CAST combines node- and token-level features using cross-attention mechanisms, surpassing previous approaches reliant on material-level embeddings like graph mean-pooling or [CLS] tokens. A masked node prediction pretraining strategy further enhances atomic-level information integration. Our method achieved up to 22.9\% improvement in property prediction across four crystal properties including band gap compared to methods like CrysMMNet and MultiMat. Pretraining was key to aligning node and text embeddings, with attention maps confirming its effectiveness in capturing relationships between nodes and tokens. This study highlights the potential of multimodal learning in materials science, paving the way for more robust predictive models that incorporate both local and global information.