Understanding the dynamics of feature learning in neural networks (NNs) remains a significant challenge. The work of Mousavi et al. analyzes a multiple index teacher–student setting and shows that a two-layer student attains a low-rank structure in its first-layer weights when trained with stochastic gradient descent (SGD) and a strong regularizer. This structural property is known to reduce sample complexity of generalization. Indeed, in a second step, the same authors establish algorithm-specific learning guarantees under additional assumptions. In this paper, we focus exclusively on the structure discovery aspect and study it under weaker assumptions, more specifically: we allow (a) NNs of arbitrary size and depth, (b) with all parameters trainable, (c) under any smooth loss function, (d) tiny regularization, and (e) trained by any method that attains a second-order stationary point (SOSP), e.g., perturbed gradient descent (PGD). At the core of our approach is a key derandomization lemma, which states that optimizing the function E_x[g_θ(W x + b)] converges to a point where W = 0, under mild conditions. The fundamental nature of this lemma directly explains structure discovery and has immediate applications in other domains including an end-to-end approximation for MAXCUT, and computing Johnson–Lindenstrauss embeddings.

Despite their proficiency in various language tasks, Large Language Models (LLMs) struggle with combinatorial problems like Satisfiability, Traveling Salesman Problem, or even basic arithmetic. We address this gap through a novel trial & error approach for solving problems in the class NP, where candidate solutions are iteratively generated and efficiently validated using verifiers. We focus on the paradigmatic task of Sudoku and achieve state-of-the-art accuracy (99%) compared to prior neuro-symbolic approaches. Unlike prior work that used custom architectures, our method employs a vanilla decoder-only Transformer (GPT-2) without external tools or function calling. Our method integrates imitation learning of simple Sudoku rules with an explicit Depth-First Search (DFS) exploration strategy involving informed guessing and backtracking. Moving beyond imitation learning, we seek to minimize the number of guesses until reaching a solution. This is achieved using depth-1 guessing, showing empirically that almost all Sudoku can be solved using the puzzle's rules with at most one guess. We provide a rigorous analysis of this setup formalizing its connection to a contextual variant of Min–Sum Set Cover, a well-studied problem in algorithms and stochastic optimization.

In an era dominated by digital interactions, phishing campaigns have evolved to exploit not just technological vulnerabilities but also human traits. This study takes an unprecedented deep dive into large-scale phishing campaigns aimed at Meta's users, offering a dual perspective on the technical mechanics and human elements involved. Analysing data from over 25,000 victims worldwide, we highlight the nuances of these campaigns, from the intricate techniques deployed by the attackers to the sentiments and behaviours of those targeted. Unlike prior research conducted in controlled environments, this investigation capitalises on the vast, diverse, and genuine data extracted directly from active phishing campaigns, allowing for a more holistic understanding of the drivers, facilitators, and human factors. Through applying advanced computational techniques, including natural language processing and machine learning, this work unveils critical insights into the psyche of victims and the evolving tactics of modern phishers. Our analysis illustrates very poor password selection choices from the victims, with 30.27% of them picking low-complexity passwords and 58.23% reusing leaked passwords. Additionally, more than 10% exhibit strong persistence in re-victimisation by posting again to the phishing platforms of the same phishers. Finally, we reveal many correlations regarding demographics and the time periods when victims are more vulnerable during the day, as well as analyse the sentiment, emotion, and tone of text responses that they submitted, illustrating how convinced they were of the scam.

The extensive surviving corpus of the ancient scholar Plutarch of Chaeronea (ca. 45-120 CE) also contains several texts which, according to current scholarly opinion, did not originate with him and are therefore attributed to an anonymous author Pseudo-Plutarch. These include, in particular, the work Placita Philosophorum (Quotations and Opinions of the Ancient Philosophers), which is extremely important for the history of ancient philosophy. Little is known about the identity of that anonymous author and its relation to other authors from the same period. This paper presents a BERT language model for Ancient Greek. The model discovers previously unknown statistical properties relevant to these literary, philosophical, and historical problems and can shed new light on this authorship question. In particular, the Placita Philosophorum, together with one of the other Pseudo-Plutarch texts, shows similarities with the texts written by authors from an Alexandrian context (2nd/3rd century CE).