Large language model (LLM) safety classifiers such as Llama Guard are effective at detecting overtly harmful prompts but remain vulnerable to adversarial jailbreak attacks that disguise malicious intent through role-play scenarios, fictional framing, and indirect requests. We present Reflect-Guard, a method that augments LLM-based safety classifiers with chain-of-thought self-reflection capabilities through parameter-efficient fine-tuning. Our approach distills analytical reasoning from GPT-4o-mini into structured reflection annotations, then trains Llama-Guard-3-8B via QLoRA to generate logical self-reflections before issuing safety verdicts. Using only 1000 training examples and updating just 0.5
This paper identifies over 100 inflation shock episodes in 56 countries since the 1970s, including over 60 episodes linked to the 1973–79 oil crises. We document that only in 60 percent of the episodes was inflation brought back down (or "resolved") within 5 years, and that even in these "successful" cases resolving inflation took, on average, over 3 years. Success rates were lower and resolution times longer for episodes induced by terms-of-trade shocks during the 1973–79 oil crises. Most unresolved episodes involved "premature celebrations", where inflation declined initially, only to plateau at an elevated level or re-accelerate. Сountries that resolved inflation had tighter monetary policy that was maintained more consistently over time, lower nominal wage growth, and less currency depreciation, compared to unresolved cases. Successful disinflations were associated with short-term output losses, but not with larger output, employment, or real wage losses over a 5-year horizon, potentially indicating the value of policy credibility and macroeconomic stability.
The emergence of large-scale Low Earth Orbit (LEO) satellite constellations has renewed attention to satellite networks, with a vision to deliver global connectivity and performance levels comparable to terrestrial infrastructures. Despite this progress, such constellations are often treated as standalone systems rather than being fully integrated into next-generation communication architectures. To bridge this gap, extensive research has been devoted to the development of Non-Terrestrial Networks (NTN), aiming to enable unified operation across terrestrial and space segments within cellular infrastructures. Nonetheless, realizing this vision remains challenging due to persistent issues in latency, dynamic scheduling, and resource management. To tackle these challenges, we present the development of a Space Cloud, where Multi-access Edge Computing (MEC) services are deployed across satellite nodes and interconnected through Inter-Satellite Links (ISL), forming a distributed space-based data center. By enabling in-orbit computation, the Space Cloud reduces dependence on terrestrial infrastructure, as processing no longer needs to be offloaded to ground-based data centers. This architectural shift is key to meeting the lowlatency requirements of modern, latency-sensitive applications. This paper proposes a distributed Reinforcement Learning (RL) approach to manage computational resources in space-based MEC environments. The system leverages a scalable actorcritic framework, where neural network-based actors and critics are deployed on individual satellites. Each node operates autonomously and in a distributed manner, enabling the network to optimize actions locally while considering both instant rewards and long-term impact. The RL model leverages historical data and processing patterns to control the activation of on-orbit servers, aiming for efficient resource utilization. We evaluate the proposed strategy using a synthetic constellation designed with our in-house satellite emulation and MEC computation framework. Moreover, through Pareto-efficient analysis across key performance indicators (KPI), we benchmark our approach against conventional baselines and assess its applicability under the constraints of specific space missions. The results demonstrate that the RL controller achieves comparable task failure and latency rates to the baselines, while significantly reducing resource consumption.
The market for AI software-development tools has expanded faster than the frameworks used to evaluate it, leaving practitioners to choose among code-completion assistants, AI-native integrated development environments (IDEs), and terminal-native agents on the basis of headline price or benchmark rank — neither of which predicts realised value. This paper makes two contributions. First, it develops the Cost–Methodology–Fit (CMF) framework, an analytical model that treats tool selection as the alignment of a team's dominant workflow with a tool's interaction paradigm and billing structure, grounded in the established SPACE model of developer productivity [1]. Second, drawing on a systematic documentary comparison of leading tools (verified June 2026) and on the conflicting experimental literature — a controlled trial reporting a 55.8% task speed-up [2] against a randomised trial of experienced developers reporting a net slowdown [3] — it derives the central claim that AI-tool value is workflow-contingent, not tool-intrinsic. Because documentary comparison cannot establish causal productivity effects, the paper additionally specifies a reproducible mixed-methods evaluation protocol that adopting organisations can run to measure fit in their own context. We report the framework and protocol, not new empirical outcomes, and state this scope explicitly. Findings indicate the market has bifurcated by billing model, that capability is increasingly a model-level rather than tool-level property, and that hybrid tool stacks are a rational response to fit-contingency. Keywords: developer productivity; SPACE framework; usage-based pricing; agentic development; evaluation protocol; tool selection
This paper presents two approaches to enhance the accuracy and robustness of the conventional estimator from [1] for affine–state nonlinear systems affected by external perturbations and measurement noise. The proposed methods are: 1) integration of additional filtering mechanisms; and 2) modification of the conventional adaptive observer using the heavy–ball algorithm. The analysis shows the input-to-state stability property with a uniform exponential convergence rate under the regressor excitation assumption. These proposals offer the following improvements compared to [2]: a) disturbances and noise have a significantly weaker effect on parameter estimation, as the new filtering processes attenuate the propagation of high–frequency perturbations; and b) the convergence and robustness properties are no longer directly governed by the excitation signal, allowing for a more controlled convergence behavior. In addition, some simulations confirm superior accuracy compared to the conventional adaptive observers.
