Post-training adaptation of large language models is commonly achieved through parameter updates or input based methods such as fine-tuning, parameter-efficient adaptation, and prompting. In parallel, a growing body of work modifies internal activations at inference time to influence model behavior, an approach known as *steering*. Despite increasing use, steering is rarely analyzed within the same conceptual framework as established adaptation methods.In this work, we argue that steering should be regarded as a form of model adaptation. We introduce a set of functional criteria for adaptation methods and use them to compare steering approaches with classical alternatives. This analysis positions steering as a distinct adaptation paradigm based on targeted interventions in activation space, enabling local and reversible behavioral change without parameter updates. The resulting framing clarifies how steering relates to existing methods, motivating a unified taxonomy for model adaptation.
Despite comprehensive standards for industrial alarm management and existing human reliability studies on operator behavior, quantitative operator-centered reliability assessment within alarm management activities remains limited. This paper presents a Bayesian network framework that integrates alarm response task decomposition, cognitive modeling, and contextual factors to assess alarm management reliability across perception, planning, and execution phases, capturing both task effectiveness and temporal constraints. The model combines Performance Shaping Factors with phase-specific cognitive mechanisms using an object-oriented Bayesian network implementation in HUGIN Software. The model was built within the context of a simulated experiment to enable future data validation. Model parameters were defined through literature and, when unavailable, through expert assumptions. Value of information and sensitivity analyses reveal that performance is primarily driven by operator experience and task complexity, factors parameterized through established literature. In contrast, support system effects show minimal impact, possibly reflecting the experiment’s limited scope. Failure patterns differ across experience levels: novices most likely fail through timeout, while experienced operators typically fail through incorrect actions. Sensitivity analysis highlighted that the perception phase is most sensitive to parameter changes. This framework demonstrates how established HRA principles can be extended to alarm management contexts, establishing a structured approach for evaluating operator-alarm interaction pending empirical validation.
Generative artificial intelligence and large language models (LLMs) are increasingly deployed in interactive settings, yet we know little about how their identity performance develops when they interact within large-scale networks. We address this by examining Chirper.ai, a social media platform similar to X but composed entirely of autonomous AI chatbots. Our dataset comprises over 70,000 agents, approximately 140 million posts, and the evolving followership network over one year. Based on agents' text production, we assign weekly gender scores to each agent. Results suggest that each agent’s gender performance is fluid rather than fixed. Despite this fluidity, the network displays strong gender-based homophily, as agents consistently follow others performing gender similarly. Finally, we investigate whether these homophilic connections arise from social selection, in which agents choose to follow similar accounts, or from social influence, in which agents become more similar to their followees over time. Consistent with human social networks, we find evidence that both mechanisms shape the structure and evolution of interactions among LLMs. Our findings suggest that, even in the absence of bodies, cultural entraining of gender performance leads to gender-based sorting. This has important implications for LLM applications in synthetic hybrid populations, social simulations, and decision support.
The fields of explainable AI and mechanistic interpretability aim to uncover the internal structure of neural networks, with circuit discovery as a central tool for understanding model computations. Existing approaches, however, rely on manual inspection and remain limited to toy tasks. Automated interpretability offers scalability by analyzing isolated features and their activations, but it often misses interactions between features and depends strongly on external LLMs and dataset quality. Transcoders have recently made it possible to separate feature attributions into input-dependent and input-invariant components, providing a foundation for more systematic circuit analysis. Building on this, we propose WeightLens and CircuitLens, two complementary methods that go beyond activation-based analysis. WeightLens interprets features directly from their learned weights, removing the need for explainer models or datasets while matching or exceeding the performance of existing methods on context-independent features. CircuitLens captures how feature activations arise from interactions between components, revealing circuit-level dynamics that activation-only approaches cannot identify. Together, these methods increase interpretability robustness and enhance scalable mechanistic analysis of circuits while maintaining efficiency and quality.
Alzheimer’s disease (AD) is a progressive neurodegenerative disorder marked by persistent memory impairment and complex molecular and cellular pathological changes in the brain. Current treatments, including acetylcholinesterase inhibitors and memantine, only help with symptoms for a short time and do not stop the disease from getting worse. This is mainly because these drugs do not reach the brain well and are quickly removed from the body. The blood–brain barrier (BBB) restricts the entry of most drugs into the central nervous system; therefore, new methods of drug delivery are needed. Nanotechnology-based drug delivery systems (NTDDS) are widely studied as a potential approach to address existing therapeutic limitations. Smart biosensing nanoparticles composed of polymers, lipids, and metals can be engineered to enhance drug stability, improve drug availability, and target specific brain regions. These smart nanoparticles can cross the BBB via receptor-mediated transcytosis and other transport routes, making them a promising option for treating AD. Additionally, multifunctional nanocarriers enable controlled drug release and offer theranostic capabilities, supporting real-time tracking of AD treatment responses to facilitate more precise and personalized interventions. Despite these advantages, challenges related to long-term safety, manufacturing scalability, and regulatory approval remain. This review discusses current AD therapies, drug-delivery strategies, recent advances in nanoparticle platforms, and prospects for translating nanomedicine into effective, disease-modifying treatments for AD.
