Comparison of fine-tuning models with reinforcement learning (RL) and supervised fine-tuning (SFT) reveals that, despite similar performance at a new task, RL preserves prior knowledge and capabilities significantly better. We find that the degree of forgetting is determined by the distributional shift, measured as the KL-divergence between the fine-tuned and base policy evaluated on the new task. Our analysis reveals that on-policy RL is implicitly biased towards KL-minimal solutions among the many that solve the new task, whereas SFT can converge to distributions arbitrarily far from the base model. We validate these findings through experiments with large language models and robotic foundation models and further provide theoretical justification for why on-policy RL updates lead to a smaller KL change. We term this principle $\textit{RL's Razor}$: among all ways to solve a new task, RL prefers those closest in KL to the original model.

Vision Language Models (VLMs) have recently achieved significant progress in bridging visual perception and linguistic reasoning. Recently, OpenAI o3 model introduced a zoom-in search strategy that effectively elicits active perception capabilities in VLMs, improving downstream task performance. However, enabling VLMs to reason effectively over appropriate image regions remains a core challenge in GUI grounding, particularly under high-resolution inputs and complex multi-element visual interactions. In this work, we propose LASER, a self-evolving framework that progressively endows VLMs with multi-step perception capabilities, enabling precise coordinate prediction. Specifically, our approach integrate Monte Carlo quality estimation with Intersection-over-Union (IoU)-based region quality evaluation to jointly encourage both accuracy and diversity in constructing high-quality preference data. This combination explicitly guides the model to focus on instruction-relevant key regions while adaptively allocating reasoning steps based on task complexity. Comprehensive experiments on the ScreenSpot Pro and ScreenSpot-v2 benchmarks demonstrate consistent performance gains, validating the effectiveness of our method. Furthermore, when fine-tuned on GT A1-7B, LASER achieves a score of 55.7 on the ScreenSpot-Pro benchmark, establishing a new state-of-the-art (SoT A) among 7B-scale models.

--This paper proposes a structured methodology to evaluate AI-generated game narratives, leveraging the Delphi study structure with a panel of narrative design experts. Our approach synthesizes story quality dimensions from literature and expert insights, mapping them into the Kano model framework to understand their impact on player satisfaction. The results can inform game developers on prioritizing quality aspects when co-creating game narratives with generative AI. While generative AI has surged into public and research consciousness following the release of systems like ChatGPT, video games have a longer tradition of using AI techniques to generate content that would otherwise be authored by human designers. This tradition is well established in the field of Procedural Content Generation, which encompasses a range of methods for algorithmically creating game elements such as levels, characters, quests, and storylines [1].

Long-range dependency is one of the most desired properties of recent sequence models such as state-space models (particularly Mamba) and transformer models. New model architectures are being actively developed and benchmarked for prediction tasks requiring long-range dependency. However, the capability of modeling long-range dependencies of these models has not been investigated from a theoretical perspective, which hinders a systematic improvement on this aspect. In this work, we mathematically define long-range dependency using the derivative of hidden states with respect to past inputs and compare the capability of SSM and transformer models of modeling long-range dependency based on this definition. We showed that the long-range dependency of SSM decays exponentially with the sequence length, which aligns with the exponential decay of memory function in RNN. But the attention mechanism used in transformers is more flexible and is not constrained to exponential decay, which could in theory perform better at modeling long-range dependency with sufficient training data, computing resources, and proper training. To combine the flexibility of long-range dependency of attention mechanism and computation efficiency of SSM, we propose a new formulation for hidden state update in SSM and prove its stability under a standard Gaussian distribution of the input data.

State estimation in control and systems engineering traditionally requires extensive manual system identification or data-collection effort. However, transformer-based foundation models in other domains have reduced data requirements by leveraging pre-trained generalist models. Ultimately, developing zero-shot foundation models of system dynamics could drastically reduce manual deployment effort. While recent work shows that transformer-based end-to-end approaches can achieve zero-shot performance on unseen systems, they are limited to sensor models seen during training. We introduce the foundation model unscented Kalman filter (FM-UKF), which combines a transformer-based model of system dynamics with analytically known sensor models via an UKF, enabling generalization across varying dynamics without retraining for new sensor configurations. We evaluate FM-UKF on a new benchmark of container ship models with complex dynamics, demonstrating a competitive accuracy, effort, and robustness trade-off compared to classical methods with approximate system knowledge and to an end-to-end approach. The benchmark and dataset are open sourced to further support future research in zero-shot state estimation via foundation models.

Chronic obesity management requires continuous monitoring of energy balance behaviors, yet traditional self-reported methods suffer from significant underreporting and recall bias, and difficulty in integration with modern digital health systems. This study presents COBRA (Chronic Obesity Behavioral Recognition Architecture), a novel deep learning framework for objective behavioral monitoring using wrist-worn multimodal sensors. COBRA integrates a hybrid D-Net architecture combining U-Net spatial modeling, multi-head self-attention mechanisms, and BiLSTM temporal processing to classify daily activities into four obesity-relevant categories: Food Intake, Physical Activity, Sedentary Behavior, and Daily Living. Validated on the WISDM-Smart dataset with 51 subjects performing 18 activities, COBRA's optimal prepro-cessing strategy combines spectral-temporal feature extraction, achieving high performance across multiple architectures. D-Net demonstrates 96.86% overall accuracy with category-specific F1-scores of 98.55% (Physical Activity), 95.53% (Food Intake), 94.63% (Sedentary Behavior), and 98.68% (Daily Living), outperforming state-of-the-art baselines by 1.18% in accuracy. The framework shows robust generalizability with low demographic variance ( < 3%), enabling scalable deployment for personalized obesity interventions and continuous lifestyle monitoring.

The emergence of large language models (LLMs) has introduced a new paradigm in automation: LLM agents or Agentic Automation with Computer Use (AACU). Unlike traditional Robotic Process Automation (RPA), which relies on rule - based workflows and scripting, AACU enables intelligent agents to perform tasks through natural language instructions and autonomous inte raction with user interfaces. This study investigates whether AACU can serve as a viable alternative to RPA in enterprise workflow automation. We conducted controlled experiments across three standard RPA challenges data entry, monitoring, and document extraction comparing RPA (via UiPath) and AACU (via Anthropic's Computer Use Agent) in terms of speed, reliability, and development effort. Results indicate that RPA outperforms AACU in execution speed and reliability, particularly in repetitive, stable environments. However, AACU significantly reduces development time and adapts more flexibly to dynamic interfaces. While current AACU implementations are not yet production - ready, their promise in rapid prototyping and lightweight automation is evident. Future research should explore multi - agent orchestration, hybrid RPA - AACU architectures, and more robust evaluation a cross industries and platforms.

The widespread adoption of Kubernetes (K8s) for orchestrating cloud-native applications has introduced significant security challenges, such as misconfigured resources and overly permissive configurations. Failing to address these issues can result in unauthorized access, privilege escalation, and lateral movement within clusters. Most existing K8s security solutions focus on detecting misconfigurations, typically through static analysis or anomaly detection. In contrast, this paper presents KubeGuard, a novel runtime log-driven recommender framework aimed at mitigating risks by addressing overly permissive configurations. KubeGuard is designed to harden K8s environments through two complementary tasks: Resource Creation and Resource Refinement. It leverages large language models (LLMs) to analyze manifests and runtime logs reflecting actual system behavior, using modular prompt-chaining workflows. This approach enables KubeGuard to create least-privilege configurations for new resources and refine existing manifests to reduce the attack surface. KubeGuard's output manifests are presented as recommendations that users (e.g., developers and operators) can review and adopt to enhance cluster security. Our evaluation demonstrates that KubeGuard effectively generates and refines K8s manifests for Roles, NetworkPolicies, and Deployments, leveraging both proprietary and open-source LLMs. The high precision, recall, and F1-scores affirm KubeGuard's practicality as a framework that translates runtime observability into actionable, least-privilege configuration guidance.

Autoregressive next token prediction language models offer powerful capabilities but face significant challenges in practical deployment due to the high computational and memory costs of inference, particularly during the decoding stage. We introduce Set Block Decoding (SBD), a simple and flexible paradigm that accelerates generation by integrating standard next token prediction (NTP) and masked token prediction (MATP) within a single architecture. SBD allows the model to sample multiple, not necessarily consecutive, future tokens in parallel, a key distinction from previous acceleration methods. This flexibility allows the use of advanced solvers from the discrete diffusion literature, offering significant speedups without sacrificing accuracy. SBD requires no architectural changes or extra training hyperparameters, maintains compatibility with exact KV-caching, and can be implemented by fine-tuning existing next token prediction models. By fine-tuning Llama-3.1 8B and Qwen-3 8B, we demonstrate that SBD enables a 3-5x reduction in the number of forward passes required for generation while achieving same performance as equivalent NTP training.

In linguistics, coherence can be achieved by different means, such as by maintaining reference to the same set of entities across sentences and by establishing discourse relations between them. However, most existing work on coherence modeling focuses exclusively on either entity features or discourse relation features, with little attention given to combining the two. In this study, we explore two methods for jointly modeling entities and discourse relations for coherence assessment. Experiments on three benchmark datasets show that integrating both types of features significantly enhances the performance of coherence models, highlighting the benefits of modeling both simultaneously for coherence evaluation.