While prompt engineering has emerged as a crucial technique for optimizing large language model performance, the underlying optimization landscape remains poorly understood. Current approaches treat prompt optimization as a black-box problem, applying sophisticated search algorithms without characterizing the landscape topology they navigate. We present a systematic analysis of fitness landscape structures in prompt engineering using autocorrelation analysis across semantic embedding spaces. Through experiments on error detection tasks with two distinct prompt generation strategies -- systematic enumeration (1,024 prompts) and novelty-driven diversification (1,000 prompts) -- we reveal fundamentally different landscape topologies. Systematic prompt generation yields smoothly decaying autocorrelation, while diversified generation exhibits non-monotonic patterns with peak correlation at intermediate semantic distances, indicating rugged, hierarchically structured landscapes. Task-specific analysis across 10 error detection categories reveals varying degrees of ruggedness across different error types. Our findings provide an empirical foundation for understanding the complexity of optimization in prompt engineering landscapes.

We introduce SasAgent, a multi-agent AI system powered by large language models (LLMs) that automates small-angle scattering (SAS) data analysis by leveraging tools from the SasView software and enables user interaction via text input. SasAgent features a coordinator agent that interprets user prompts and delegates tasks to three specialized agents for scattering length density (SLD) calculation, synthetic data generation, and experimental data fitting. These agents utilize LLM-friendly tools to execute tasks efficiently. These tools, including the model data tool, Retrieval-Augmented Generation (RAG) documentation tool, bump fitting tool, and SLD calculator tool, are derived from the SasView Python library. A user-friendly Gradio-based interface enhances user accessibility. Through diverse examples, we demonstrate SasAgent's ability to interpret complex prompts, calculate SLDs, generate accurate scattering data, and fit experimental datasets with high precision. This work showcases the potential of LLM-driven AI systems to streamline scientific workflows and enhance automation in SAS research.

Large Language Models (LLMs) have demonstrated effectiveness across a wide variety of tasks involving natural language, however, a fundamental problem of hallucinations still plagues these models, limiting their trustworthiness in generating consistent, truthful information. Detecting hallucinations has quickly become an important topic, with various methods such as uncertainty estimation, LLM Judges, retrieval augmented generation (RAG), and consistency checks showing promise. Many of these methods build upon foundational metrics, such as ROUGE, BERTScore, or Perplexity, which often lack the semantic depth necessary to detect hallucinations effectively. In this work, we propose a novel approach inspired by ROUGE that constructs an N-Gram frequency tensor from LLM-generated text. This tensor captures richer semantic structure by encoding co-occurrence patterns, enabling better differentiation between factual and hallucinated content. We demonstrate this by applying tensor decomposition methods to extract singular values from each mode and use these as input features to train a multi-layer perceptron (MLP) binary classifier for hallucinations. Our method is evaluated on the HaluEval dataset and demonstrates significant improvements over traditional baselines, as well as competitive performance against state-of-the-art LLM judges.

Traditional fixed architectures struggle to cope with dynamic and unpredictable environments, leading to inefficiencies in energy consumption and connectivity. This paper addresses this gap by proposing an adaptive architecture for drone swarms, leveraging a Large Language Model (LLM) to dynamically select the optimal architecture--centralized, hierarchical, or holonic--based on real-time mission parameters such as task complexity, swarm size, and communication stability. Our system addresses the challenges of scalability, adaptability, and robustness, ensuring efficient energy consumption and maintaining connectivity under varying conditions. Extensive simulations demonstrate that our adaptive architecture outperforms traditional static models in terms of scalability, energy efficiency, and connectivity. These results highlight the potential of our approach to provide a scalable, adaptable, and resilient solution for real-world disaster response scenarios.

We introduce Plantbot, a hybrid lifeform that connects a living plant with a mobile robot through a network of large language model (LLM) modules. Each module - responsible for sensing, vision, dialogue, or action - operates asynchronously and communicates via natural language, enabling seamless interaction across biological and artificial domains. This architecture leverages the capacity of LLMs to serve as hybrid interfaces, where natural language functions as a universal protocol, translating multimodal data (soil moisture, temperature, visual context) into linguistic messages that coordinate system behaviors. The integrated network transforms plant states into robotic actions, installing normativity essential for agency within the sensor-motor loop. By combining biological and robotic elements through LLM-mediated communication, Plantbot behaves as an embodied, adaptive agent capable of responding autonomously to environmental conditions. This approach suggests possibilities for a new model of artificial life, where decentralized, LLM modules coordination enable novel interactions between biological and artificial systems.

Real world deployments often expose modern object recognition models to domain shifts that precipitate a severe drop in accuracy. Such shifts encompass (i) variations in low level image statistics, (ii) changes in object pose and viewpoint, (iii) partial occlusion, and (iv) visual confusion across adjacent classes. To mitigate this degradation, we introduce the Re-Thinking Vision Language Model (RT -VLM) framework. The foundation of this framework is a unique synthetic dataset generation pipeline that produces images annotated with "4-Clues": precise bounding boxes, class names, detailed object-level captions, and a comprehensive context-level caption for the entire scene. We then perform parameter efficient supervised tuning of Llama 3.2 11B Vision Instruct on this resource. At inference time, a two stage Re-Thinking scheme is executed: the model first emits its own four clues, then re examines these responses as evidence and iteratively corrects them. Across robustness benchmarks that isolate individual domain shifts, RT -VLM consistently surpasses strong baselines. These findings indicate that the integration of structured multimodal evidence with an explicit self critique loop constitutes a promising route toward reliable and transferable visual understanding.

The growing complexity of cyber incidents presents significant challenges for digital forensic investigators, especially in evidence collection and analysis. Public resources are still limited because of ethical, legal, and privacy concerns, even though realistic datasets are necessary to support research and tool developments. To address this gap, we introduce ForensicsData, an extensive Question-Context-Answer (Q-C-A) dataset sourced from actual malware analysis reports. It consists of more than 5,000 Q-C-A triplets. A unique workflow was used to create the dataset, which extracts structured data, uses large language models (LLMs) to transform it into Q-C-A format, and then uses a specialized evaluation process to confirm its quality. Among the models evaluated, Gemini 2 Flash demonstrated the best performance in aligning generated content with forensic terminology. ForensicsData aims to advance digital forensics by enabling reproducible experiments and fostering collaboration within the research community.

The integration of Large Language Models (LLMs) in 6G vehicular networks promises unprecedented advancements in intelligent transportation systems. However, offloading LLM computations from vehicles to edge infrastructure poses significant privacy risks, potentially exposing sensitive user data. This paper presents a novel privacy-preserving offloading framework for LLM-integrated vehicular networks. We introduce a hybrid approach combining federated learning (FL) and differential privacy (DP) techniques to protect user data while maintaining LLM performance. Our framework includes a privacy-aware task partitioning algorithm that optimizes the trade-off between local and edge computation, considering both privacy constraints and system efficiency. We also propose a secure communication protocol for transmitting model updates and aggregating results across the network. Experimental results demonstrate that our approach achieves 75\% global accuracy with only a 2-3\% reduction compared to non-privacy-preserving methods, while maintaining DP guarantees with an optimal privacy budget of $\varepsilon = 0.8$. The framework shows stable communication overhead of approximately 2.1MB per round with computation comprising over 90\% of total processing time, validating its efficiency for resource-constrained vehicular environments.

The use of unvetted third-party and internet data renders pre-trained models susceptible to backdoor attacks. Detecting backdoor samples is critical to prevent backdoor activation during inference or injection during training. However, existing detection methods often require the defender to have access to the poisoned models, extra clean samples, or significant computational resources to detect backdoor samples, limiting their practicality. To address this limitation, we propose a backdoor sample detection method based on perturbatio\textbf{N} discr\textbf{E}pancy consis\textbf{T}ency \textbf{E}valuation (\NETE). This is a novel detection method that can be used both pre-training and post-training phases. In the detection process, it only requires an off-the-shelf pre-trained model to compute the log probability of samples and an automated function based on a mask-filling strategy to generate perturbations. Our method is based on the interesting phenomenon that the change in perturbation discrepancy for backdoor samples is smaller than that for clean samples. Based on this phenomenon, we use curvature to measure the discrepancy in log probabilities between different perturbed samples and input samples, thereby evaluating the consistency of the perturbation discrepancy to determine whether the input sample is a backdoor sample. Experiments conducted on four typical backdoor attacks and five types of large language model backdoor attacks demonstrate that our detection strategy outperforms existing zero-shot black-box detection methods.

A conventional LLM Unlearning setting consists of two subsets -"forget" and "retain", with the objectives of removing the undesired knowledge from the forget set while preserving the remaining knowledge from the retain. In privacy-focused unlearning research, a retain set is often further divided into neighbor sets, containing either directly or indirectly connected to the forget targets; and augmented by a general-knowledge set. A common practice in existing benchmarks is to employ only a single neighbor set, with general knowledge which fails to reflect the real-world data complexities and relationships. LLM Unlearning typically involves 1:1 sampling or cyclic iteration sampling. However, the efficacy and stability of these de facto standards have not been critically examined. In this study, we systematically evaluate these common practices. Our findings reveal that relying on a single neighbor set is suboptimal and that a standard sampling approach can obscure performance trade-offs. Based on this analysis, we propose and validate an initial set of best practices: (1) Incorporation of diverse neighbor sets to balance forget efficacy and model utility, (2) Standard 1:1 sampling methods are inefficient and yield poor results, (3) Our proposed Modular Entity-Level Unlearning (MELU) strategy as an alternative to cyclic sampling. We demonstrate that this modular approach, combined with robust algorithms, provides a clear and stable path towards effective unlearning.