Modern language model (LM) training has been divided into multiple stages, making it difficult for downstream developers to evaluate the impact of design choices made at each stage. We present EvoLM, a model suite that enables systematic and transparent analysis of LMs' training dynamics across pre-training, continued pre-training, supervised fine-tuning, and reinforcement learning. We train over 100 LMs with 1B and 4B parameters from scratch, and evaluate both upstream (language modeling) and downstream (problem-solving) capabilities, including considerations of both in-domain and out-of-domain generalization. Key insights highlight the diminishing returns from excessive pre-training and post-training, the importance and practices of mitigating forgetting during domain-specific continued pre-training, the crucial role of continued pre-training in bridging pre-training and post-training phases, and various intricate trade-offs when configuring supervised fine-tuning and reinforcement learning. To facilitate open research and reproducibility, we release all pre-trained and post-trained models, training datasets for all stages, and our entire training and evaluation pipeline.

The ongoing evolution of language models has led to the development of large-scale architectures that demonstrate exceptional performance across a wide range of tasks. However, these models come with significant computational and energy demands, as well as potential privacy implications. In this context, Small Reasoning Language Models (SRLMs) with approximately 0.5 billion parameters present a compelling alternative due to their remarkable computational efficiency and cost-effectiveness, particularly in resource-constrained environments. Despite these advantages, the limited capacity of 0.5 billion parameter models poses challenges in handling complex tasks such as mathematical reasoning. This research investigates various training strategies, including supervised fine-tuning (SFT), knowledge distillation (KD), and reinforcement learning (RL), as well as their hybrid implementations, to enhance the performance of 0.5B SRLMs. We analyze effective methodologies to bridge the performance gap between SRLMS and larger models and present insights into optimal training pipelines tailored for these smaller architectures. Through extensive experimental validation and analysis, our work aims to provide actionable recommendations for maximizing the reasoning capabilities of 0.5B models.

Deep reinforcement learning (DRL) has achieved remarkable success in a wide range of sequential decision-making applications, including robotics, healthcare, smart grids, and finance. Recent studies reveal that adversaries can implant backdoors into DRL agents during the training phase. These backdoors can later be activated by specific triggers during deployment, compelling the agent to execute targeted actions and potentially leading to severe consequences, such as drone crashes or vehicle collisions. However, existing backdoor attacks utilize simplistic and heuristic trigger configurations, overlooking the critical impact of trigger design on attack effectiveness. To address this gap, we introduce TooBadRL, the first framework to systematically optimize DRL backdoor triggers across three critical aspects: injection timing, trigger dimension, and manipulation magnitude. Specifically, we first introduce a performance-aware adaptive freezing mechanism to determine the injection timing during training. Then, we formulate trigger selection as an influence attribution problem and apply Shapley value analysis to identify the most influential trigger dimension for injection. Furthermore, we propose an adversarial input synthesis method to optimize the manipulation magnitude under environmental constraints. Extensive evaluations on three DRL algorithms and nine benchmark tasks demonstrate that TooBadRL outperforms five baseline methods in terms of attack success rate while only slightly affecting normal task performance. We further evaluate potential defense strategies from detection and mitigation perspectives. We open-source our code to facilitate reproducibility and further research.

Foundation models (FMs) are changing the way medical images are analyzed by learning from large collections of unlabeled data. Instead of relying on manually annotated examples, FMs are pre-trained to learn general-purpose visual features that can later be adapted to specific clinical tasks with little additional supervision. In this review, we examine how FMs are being developed and applied in pathology, radiology, and ophthalmology, drawing on evidence from over 150 studies. We explain the core components of FM pipelines, including model architectures, self-supervised learning methods, and strategies for downstream adaptation. We also review how FMs are being used in each imaging domain and compare design choices across applications. Finally, we discuss key challenges and open questions to guide future research.

Unmanned Aerial Vehicles (UAVs) have become essential tools in inspection and emergency response operations due to their high maneuverability and ability to access hard-to-reach areas. However, their limited battery life significantly restricts their use in long-duration missions. This paper presents a tethered marsupial robotic system composed of a UAV and an Unmanned Ground Vehicle (UGV), specifically designed for autonomous, long-duration inspection tasks in Global Navigation Satellite System (GNSS)-denied environments. The system extends the UAV's operational time by supplying power through a tether connected to high-capacity battery packs carried by the UGV. Our work details the hardware architecture based on off-the-shelf components to ensure replicability and describes our full-stack software framework used by the system, which is composed of open-source components and built upon the Robot Operating System (ROS). The proposed software architecture enables precise localization using a Direct LiDAR Localization (DLL) method and ensures safe path planning and coordinated trajectory tracking for the integrated UGV-tether-UAV system. We validate the system through three sets of field experiments involving (i) three manual flight endurance tests to estimate the operational duration, (ii) three experiments for validating the localization and the trajectory tracking systems, and (iii) three executions of an inspection mission to demonstrate autonomous inspection capabilities. The results of the experiments confirm the robustness and autonomy of the system in GNSS-denied environments. Finally, all experimental data have been made publicly available to support reproducibility and to serve as a common open dataset for benchmarking.

--Recent advancements in Large Language Models (LLMs) have significantly expanded the capabilities of artificial intelligence in natural language processing tasks. Despite this progress, their performance in specialized domains such as computer science remains relatively unexplored. Understanding the proficiency of LLMs in these domains is critical for evaluating their practical utility and guiding future developments. The POSCOMP, a prestigious Brazilian examination used for graduate admissions in computer science promoted by the Brazlian Computer Society (SBC), provides a challenging benchmark. This study investigates whether LLMs can match or surpass human performance on the POSCOMP exam. Four LLMs - ChatGPT -4, Gemini 1.0 Advanced, Claude 3 Sonnet, and Le Chat Mistral Large - were initially evaluated on the 2022 and 2023 POSCOMP exams. The assessments measured the models' proficiency in handling complex questions typical of the exam. LLM performance was notably better on text-based questions than on image interpretation tasks. In the 2022 exam, ChatGPT - 4 led with 57 correct answers out of 69 questions, followed by Gemini 1.0 Advanced (49), Le Chat Mistral (48), and Claude 3 Sonnet (44). Similar trends were observed in the 2023 exam. ChatGPT -4 achieved the highest performance, surpassing all students who took the POSCOMP 2023 exam. LLMs, particularly ChatGPT -4, show promise in text-based tasks on the POSCOMP exam, although image interpretation remains a challenge. Given the rapid evolution of LLMs, we expanded our analysis to include more recent models - o1, Gemini 2.5 Pro, Claude 3.7 Sonnet, and o3-mini-high - evaluated on the 2022-2024 POSCOMP exams. These newer models demonstrate further improvements and consistently surpass both the average and top-performing human participants across all three years. The POSCOMP [1] is a prestigious assessment designed to test the knowledge of prospective computer science graduate students, promoted by the Brazilian Computer Society (SBC). It serves as an entry criterion for many graduate programs across Brazil. Using this exam as a benchmark for evaluating Large Language Models (LLMs) allows for a direct comparison between AI capabilities and human standards, offering valuable insights into the strengths and limitations of current AI models. Recent advancements in LLMs [2], [3] have significantly expanded the capabilities of Artificial Intelligence (AI), particularly in natural language processing tasks.

Large language models (LLMs) power many modern applications, but serving them at scale remains costly and resource-intensive. Current server-centric systems overlook consumer-grade GPUs at the edge. We introduce SpecEdge, an edge-assisted inference framework that splits LLM workloads between edge and server GPUs using a speculative decoding scheme, exchanging only token outputs over the network. SpecEdge employs proactive edge drafting to overlap edge token creation with server verification and pipeline-aware scheduling that interleaves multiple user requests to increase server-side throughput. Experiments show SpecEdge enhances overall cost efficiency by 1.91x through achieving 2.22x server throughput, and reduces inter token latency by 11.24% compared to a server-only baseline, introducing a scalable, cost-effective paradigm for LLM serving. The code is available at https://github.com/kaist-ina/specedge

As AI systems become more capable, deceptive behaviors can undermine evaluation and mislead users at deployment. Recent work has shown that lie detectors can accurately classify deceptive behavior, but they are not typically used in the training pipeline due to concerns around contamination and objective hacking. We examine these concerns by incorporating a lie detector into the labelling step of LLM post-training and evaluating whether the learned policy is genuinely more honest, or instead learns to fool the lie detector while remaining deceptive. Using DolusChat, a novel 65k-example dataset with paired truthful/deceptive responses, we identify three key factors that determine the honesty of learned policies: amount of exploration during preference learning, lie detector accuracy, and KL regularization strength. We find that preference learning with lie detectors and GRPO can lead to policies which evade lie detectors, with deception rates of over 85\%. However, if the lie detector true positive rate (TPR) or KL regularization is sufficiently high, GRPO learns honest policies. In contrast, off-policy algorithms (DPO) consistently lead to deception rates under 25\% for realistic TPRs. Our results illustrate a more complex picture than previously assumed: depending on the context, lie-detector-enhanced training can be a powerful tool for scalable oversight, or a counterproductive method encouraging undetectable misalignment.

Recent experiments have shown that training trajectories of multiple deep neural networks with different architectures, optimization algorithms, hyper-parameter settings, and regularization methods evolve on a remarkably low-dimensional "hyper-ribbon-like" manifold in the space of probability distributions. Inspired by the similarities in the training trajectories of deep networks and linear networks, we analytically characterize this phenomenon for the latter. We show, using tools in dynamical systems theory, that the geometry of this low-dimensional manifold is controlled by (i) the decay rate of the eigenvalues of the input correlation matrix of the training data, (ii) the relative scale of the ground-truth output to the weights at the beginning of training, and (iii) the number of steps of gradient descent. By analytically computing and bounding the contributions of these quantities, we characterize phase boundaries of the region where hyper-ribbons are to be expected. We also extend our analysis to kernel machines and linear models that are trained with stochastic gradient descent.

Abstract--Mobile jammers pose a critical threat to 5G networks, particularly in military communications. This paper investigates an anti-jamming framework that enhances a strong adaptive beamforming baseline comprising Multiple Signal Classification (MUSIC) for Direction-of-Arrival (DoA) estimation and Minimum V ariance Distortionless Response (MVDR) for interference suppression with a lightweight machine learning (ML) model for predictive error correction. Extensive simulations in a realistic highway scenario demonstrate that the integrated system achieves a high DoA estimation accuracy of up to 99.8% and an average Signal-to-Noise Ratio (SNR) improvement of 9.58 dB. Analysis reveals that the MUSIC-MVDR baseline alone accounts for the vast majority of this performance gain (9.46 dB), indicating that the primary benefit of the simple ML model lies in correcting outlier estimates rather than providing a substantial systemic SNR increase. The framework's computational efficiency validates the effectiveness of the core beamforming approach and highlights the critical trade-off between ML model complexity and practical performance gains for securing 5G communications in contested environments.