In this paper, we introduce a combination of novel and exciting tasks: the solution and generation of linguistic puzzles. We focus on puzzles used in Linguistic Olympiads for high school students. We first extend the existing benchmark for the task of solving linguistic puzzles. We explore the use of Large Language Models (LLMs), including recent state-of-the-art models such as OpenAI's o1, for solving linguistic puzzles, analyzing their performance across various linguistic topics. We demonstrate that LLMs outperform humans on most puzzles types, except for those centered on writing systems, and for the understudied languages. We use the insights from puzzle-solving experiments to direct the novel task of puzzle generation. We believe that automating puzzle generation, even for relatively simple puzzles, holds promise for expanding interest in linguistics and introducing the field to a broader audience. This finding highlights the importance of linguistic puzzle generation as a research task: such puzzles can not only promote linguistics but also support the dissemination of knowledge about rare and understudied languages.

Sharpness-Aware Minimization (SAM) is a widely used method that steers training toward flatter minimizers, which typically generalize better. In this work, however, we show that SAM can converge to hallucinated minimizers -- points that are not minimizers of the original objective. We theoretically prove the existence of such hallucinated minimizers and establish conditions for local convergence to them. We further provide empirical evidence demonstrating that SAM can indeed converge to these points in practice. Finally, we propose a simple yet effective remedy for avoiding hallucinated minimizers.

Predicting material properties is crucial for designing better batteries, semiconductors, and medical devices. Deep learning helps scientists quickly find promising materials by predicting their energy, forces, and stresses. Companies scale capacities of deep learning models in multiple domains, such as language modeling, and invest many millions of dollars into such models. Our team analyzes how scaling training data (giving models more information to learn from), model sizes (giving models more capacity to learn patterns), and compute (giving models more computational resources) for neural networks affects their performance for material property prediction. In particular, we trained both transformer and EquiformerV2 neural networks to predict material properties. We find empirical scaling laws for these models: we can predict how increasing each of the three hyperparameters (training data, model size, and compute) affects predictive performance. In particular, the loss $L$ can be measured with a power law relationship $L = α\cdot N^{-β}$, where $α$ and $β$ are constants while $N$ is the relevant hyperparameter. We also incorporate command-line arguments for changing training settings such as the amount of epochs, maximum learning rate, and whether mixed precision is enabled. Future work could entail further investigating scaling laws for other neural network models in this domain, such as GemNet and fully connected networks, to assess how they compare to the models we trained.

Accurately forecasting chaotic systems, prevalent in domains such as weather prediction and fluid dynamics, remains a significant scientific challenge. The inherent sensitivity of these systems to initial conditions, coupled with a scarcity of observational data, severely constrains traditional modeling approaches. Since these models are typically trained for a specific system, they lack the generalization capacity necessary for real-world applications, which demand robust zero-shot or few-shot forecasting on novel or data-limited scenarios. To overcome this generalization barrier, we propose ChaosNexus, a foundation model pre-trained on a diverse corpus of chaotic dynamics. ChaosNexus employs a novel multi-scale architecture named ScaleFormer augmented with Mixture-of-Experts layers, to capture both universal patterns and system-specific behaviors. The model demonstrates state-of-the-art zero-shot generalization across both synthetic and real-world benchmarks. On a large-scale testbed comprising over 9,000 synthetic chaotic systems, it improves the fidelity of long-term attractor statistics by more than 40% compared to the leading baseline. This robust performance extends to real-world applications with exceptional data efficiency. For instance, in 5-day global weather forecasting, ChaosNexus achieves a competitive zero-shot mean error below 1 degree, a result that further improves with few-shot fine-tuning. Moreover, experiments on the scaling behavior of ChaosNexus provide a guiding principle for scientific foundation models: cross-system generalization stems from the diversity of training systems, rather than sheer data volume.

Background: Fatigue detection through physiological signals has gained growing relevance across safety-critical domains such as transportation, healthcare, and human performance monitoring. While many studies focus on individual modalities (e.g., EEG or ECG), limited attention has been given to investigating statistical relationships between signal pairs as a means to enhance classification robustness. This study aims to explore how inter-signal statistical features correlation, cross-correlation, and covariance across multiple physiological signals can support fatigue state prediction. Methodology: Using the DROZY dataset, we extracted pairwise statistical features from four physiological signals: ECG, EMG, EOG, and EEG. Fifteen distinct signal combinations were evaluated, covering uni-modal to multi-modal configurations. Feature extraction emphasized statistical relationships between signals rather than raw amplitude characteristics. The extracted features were fed into four supervised machine learning classifiers: Decision Tree (DT), Random Forest (RF), Logistic Regression (LR), and XGBoost (XGB). Performance was assessed using accuracy, precision, recall, and area under the curve (AUC). Additionally, SHAP (SHapley Additive exPlanations) values were computed to evaluate feature importance and interpret model behavior. Results: Among all classifiers and signal combinations, XGBoost applied to the EMG| EEG combination achieved the highest classification performance, with an accuracy of 0.888 and an AUC of 0.975. SHAP-based ranking revealed that the correlation between ECG and EOG-H was the most influential feature across models. Feature interaction plots indicated non-linear relationships between statistical measures and fatigue levels. The multi-signal approach consistently outperformed single-signal models, with combinations involving EEG and EMG contributing most significantly to predictive power.

Group relative policy optimization (GRPO) has demonstrated significant potential in improving the reasoning capabilities of large language models (LLMs) via reinforcement learning. However, its practical deployment is impeded by an excessively slow training process, primarily attributed to the computationally intensive autoregressive generation of multiple responses per query, which makes the generation phase the primary performance bottleneck. Although speculative decoding presents a promising direction for acceleration, its direct application in GRPO achieves limited speedup under high-concurrency training conditions. To overcome this limitation, we propose a concurrency-aware speculative decoding framework that dynamically adjusts the drafting and verification strategy according to real-time concurrency levels, thereby maximizing the acceleration of the generation process. Furthermore, to address performance degradation arising from distributional drift between the evolving target model and the fixed draft model during training, we introduce an online draft learning mechanism that enables the draft model to continuously adapt using feedback signals from the target model. Experimental results across multiple mathematical reasoning datasets and models demonstrate that the proposed method achieves end-to-end speedups of 2.35x to 2.72x, significantly surpassing baseline approaches in efficiency. The code is available at https://github.com/yedaotian9/GRPO Group relative policy optimization (GRPO) has recently emerged as a promising framework for enhancing the reasoning capabilities of large language models (LLMs) through reinforcement learning Team (2025a). In each training iteration, the LLM generates a group of responses to a given query. These responses are subsequently evaluated using a predefined rule-based reward function, and the resulting rewards are standardized prior to model updates via policy optimization Shao et al. (2024).

The dominant retrieve-then-rank pipeline in large-scale recommender systems suffers from mis-calibration and engineering overhead due to its architectural split and differing optimization objectives. While recent generative sequence models have shown promise in unifying retrieval and ranking by auto-regressively generating ranked items, existing solutions typically address either personalized search or query-free recommendation, often exhibiting performance trade-offs when attempting to unify both. We introduce SynerGen, a novel generative recommender model that bridges this critical gap by providing a single generative backbone for both personalized search and recommendation, while simultaneously excelling at retrieval and ranking tasks. Trained on behavioral sequences, our decoder-only Transformer leverages joint optimization with InfoNCE for retrieval and a hybrid pointwise-pairwise loss for ranking, allowing semantic signals from search to improve recommendation and vice versa. We also propose a novel time-aware rotary positional embedding to effectively incorporate time information into the attention mechanism. SynerGen achieves significant improvements on widely adopted recommendation and search benchmarks compared to strong generative recom-mender and joint search and recommendation baselines. This work demonstrates the viability of a single generative foundation model for industrial-scale unified information access. Large-scale search and recommendation systems in e-commerce, short video, and food-delivery platforms are typically deployed as multi-stage cascades.

Four-dimensional variational data assimilation (4DVAR) is a cornerstone of numerical weather prediction, but its cost function is difficult to optimize and computationally intensive. We propose a neural field-based reformulation in which the full spatiotemporal state is represented as a continuous function parameterized by a neural network. This reparameterization removes the time-sequential dependency of classical 4DVAR, enabling parallel-in-time optimization in parameter space. Physical constraints are incorporated directly through a physics-informed loss, simplifying implementation and reducing computational cost. We evaluate the method on the two-dimensional incompressible Navier--Stokes equations with Kolmogorov forcing. Compared to a baseline 4DVAR implementation, the neural reparameterized variants produce more stable initial condition estimates without spurious oscillations. Notably, unlike most machine learning-based approaches, our framework does not require access to ground-truth states or reanalysis data, broadening its applicability to settings with limited reference information.

The goal of coreset selection is to identify representative subsets of datasets for efficient model training. Yet, existing approaches paradoxically require expensive training-based signals, e.g., gradients, decision boundary estimates or forgetting counts, computed over the entire dataset prior to pruning, which undermines their very purpose by requiring training on samples they aim to avoid. We introduce SubZeroCore, a novel, training-free coreset selection method that integrates submodular coverage and density into a single, unified objective. To achieve this, we introduce a sampling strategy based on a closed-form solution to optimally balance these objectives, guided by a single hyperparameter that explicitly controls the desired coverage for local density measures. Despite no training, extensive evaluations show that SubZeroCore matches training-based baselines and significantly outperforms them at high pruning rates, while dramatically reducing computational overhead. SubZeroCore also demonstrates superior robustness to label noise, highlighting its practical effectiveness and scalability for real-world scenarios.

Large reasoning models improve accuracy by producing long reasoning traces, but this inflates latency and cost, motivating inference-time efficiency. We propose Retrieval-of-Thought (RoT), which reuses prior reasoning as composable "thought" steps to guide new problems. RoT organizes steps into a thought graph with sequential and semantic edges to enable fast retrieval and flexible recombination. At inference, RoT retrieves query-relevant nodes and applies reward-guided traversal to assemble a problem-specific template that guides generation. This dynamic template reuse reduces redundant exploration and, therefore, reduces output tokens while preserving accuracy. We evaluate RoT on reasoning benchmarks with multiple models, measuring accuracy, token usage, latency, and memory overhead. Findings show small prompt growth but substantial efficiency gains, with RoT reducing output tokens by up to 40%, inference latency by 82%, and cost by 59% while maintaining accuracy. RoT establishes a scalable paradigm for efficient LRM reasoning via dynamic template construction through retrieval. Large Reasoning Models (LRMs) have demonstrated impressive capabilities in solving complex tasks by producing outputs accompanied by detailed reasoning trajectories (Xu et al., 2025a). These models adopt an intentionally slower and more deliberative inference process, mimicking human-like reasoning. This approach typically involves generating longer outputs and consuming increased inference-time compute to effectively address reasoning-intensive queries. Recent efforts to improve reasoning in LLMs have primarily focused on generating more output tokens to simulate thoughtful, multi-step reasoning (Snell et al., 2024). A common approach involves guiding generation using external reward models Zhang et al. (2024). These include outcome-based reward models, such as Best-of-N (BoN) sampling.