Interactive multimodal applications (IMAs), such as route planning in the Internet of Vehicles, enrich users' personalized experiences by integrating various forms of data over wireless networks. Recent advances in large language models (LLMs) utilize mixture-of-experts (MoE) mechanisms to empower multiple IMAs, with each LLM trained individually for a specific task that presents different business workflows. In contrast to existing approaches that rely on multiple LLMs for IMAs, this paper presents a novel paradigm that accomplishes various IMAs using a single compositional LLM over wireless networks. The two primary challenges include 1) guiding a single LLM to adapt to diverse IMA objectives and 2) ensuring the flexibility and efficiency of the LLM in resource-constrained mobile environments. To tackle the first challenge, we propose ContextLoRA, a novel method that guides an LLM to learn the rich structured context among IMAs by constructing a task dependency graph. We partition the learnable parameter matrix of neural layers for each IMA to facilitate LLM composition. Then, we develop a step-by-step fine-tuning procedure guided by task relations, including training, freezing, and masking phases. This allows the LLM to learn to reason among tasks for better adaptation, capturing the latent dependencies between tasks. For the second challenge, we introduce ContextGear, a scheduling strategy to optimize the training procedure of ContextLoRA, aiming to minimize computational and communication costs through a strategic grouping mechanism. Experiments on three benchmarks show the superiority of the proposed ContextLoRA and ContextGear. Furthermore, we prototype our proposed paradigm on a real-world wireless testbed, demonstrating its practical applicability for various IMAs. We will release our code to the community.

Next generation Radio Access Networks (RANs) introduce programmability, intelligence, and near real-time control through intelligent controllers, enabling enhanced security within the RAN and across broader 5G/6G infrastructures. This paper presents a comprehensive survey highlighting opportunities, challenges, and research gaps for Large Language Models (LLMs)-assisted explainable (XAI) intrusion detection (IDS) for secure future RAN environments. Motivated by this, we propose an LLM interpretable anomaly-based detection system for distributed denial-of-service (DDoS) attacks using multivariate time series key performance measures (KPMs), extracted from E2 nodes, within the Near Real-Time RAN Intelligent Controller (Near-RT RIC). An LSTM-based model is trained to identify malicious User Equipment (UE) behavior based on these KPMs. To enhance transparency, we apply post-hoc local explainability methods such as LIME and SHAP to interpret individual predictions. Furthermore, LLMs are employed to convert technical explanations into natural-language insights accessible to non-expert users. Experimental results on real 5G network KPMs demonstrate that our framework achieves high detection accuracy (F1-score > 0.96) while delivering actionable and interpretable outputs.

LLMs for clinical decision support often fail under small but clinically meaningful input shifts such as masking a symptom or negating a finding, despite high performance on static benchmarks. These reasoning failures frequently go undetected by standard NLP metrics, which are insensitive to latent representation shifts that drive diagnosis instability. We propose a geometry-aware evaluation framework, LAPD (Latent Agentic Perturbation Diagnostics), which systematically probes the latent robustness of clinical LLMs under structured adversarial edits. Within this framework, we introduce Latent Diagnosis Flip Rate (LDFR), a model-agnostic diagnostic signal that captures representational instability when embeddings cross decision boundaries in PCA-reduced latent space. Clinical notes are generated using a structured prompting pipeline grounded in diagnostic reasoning, then perturbed along four axes: masking, negation, synonym replacement, and numeric variation to simulate common ambiguities and omissions. We compute LDFR across both foundation and clinical LLMs, finding that latent fragility emerges even under minimal surface-level changes. Finally, we validate our findings on 90 real clinical notes from the DiReCT benchmark (MIMIC-IV), confirming the generalizability of LDFR beyond synthetic settings. Our results reveal a persistent gap between surface robustness and semantic stability, underscoring the importance of geometry-aware auditing in safety-critical clinical AI.

Transformers have profoundly influenced AI research, but explaining their decisions remains challenging -- even for relatively simpler tasks such as classification -- which hinders trust and safe deployment in real-world applications. Although activation-based attribution methods effectively explain transformer-based text classification models, our findings reveal that these methods can be undermined by class-irrelevant features within activations, leading to less reliable interpretations. To address this limitation, we propose Contrast-CAT, a novel activation contrast-based attribution method that refines token-level attributions by filtering out class-irrelevant features. By contrasting the activations of an input sequence with reference activations, Contrast-CAT generates clearer and more faithful attribution maps. Experimental results across various datasets and models confirm that Contrast-CAT consistently outperforms state-of-the-art methods. Notably, under the MoRF setting, it achieves average improvements of x1.30 in AOPC and x2.25 in LOdds over the most competing methods, demonstrating its effectiveness in enhancing interpretability for transformer-based text classification.

Open-source Large Language Models (LLMs) often employ safety alignment methods to resist harmful instructions. However, recent research shows that maliciously fine-tuning these LLMs on harmful data can easily bypass these safeguards. To counter this, we theoretically uncover why malicious fine-tuning succeeds and identify potential defense strategies. Building on the theoretical analysis, we introduce the Self-Degraded Defense (SDD) framework. SDD encourages LLMs to produce high-quality but irrelevant responses to harmful prompts. When attackers attempt malicious fine-tuning, the general capability of the LLM aligned by SDD will significantly decrease, rendering it incapable of following harmful instructions. Our experimental results confirm SDD's effectiveness against such attacks.

Effectively leveraging diversity has been shown to improve performance for various machine learning models, including large language models (LLMs). However, determining the most effective way of using diversity remains a challenge. In this work, we compare two diversity approaches for answering binary questions using LLMs: model diversity, which relies on multiple models answering the same question, and question interpretation diversity, which relies on using the same model to answer the same question framed in different ways. For both cases, we apply majority voting as the ensemble consensus heuristic to determine the final answer. Our experiments on boolq, strategyqa, and pubmedqa show that question interpretation diversity consistently leads to better ensemble accuracy compared to model diversity. Furthermore, our analysis of GPT and LLaMa shows that model diversity typically produces results between the best and the worst ensemble members without clear improvement.

Pedestrian intention prediction is essential for autonomous driving in complex urban environments. Conventional approaches depend on supervised learning over frame sequences and require extensive retraining to adapt to new scenarios. Here, we introduce BF-PIP (Beyond Frames Pedestrian Intention Prediction), a zero-shot approach built upon Gemini 2.5 Pro. It infers crossing intentions directly from short, continuous video clips enriched with structured JAAD metadata. In contrast to GPT-4V based methods that operate on discrete frames, BF-PIP processes uninterrupted temporal clips. It also incorporates bounding-box annotations and ego-vehicle speed via specialized multimodal prompts. Without any additional training, BF-PIP achieves 73% prediction accuracy, outperforming a GPT-4V baseline by 18 %. These findings illustrate that combining temporal video inputs with contextual cues enhances spatiotemporal perception and improves intent inference under ambiguous conditions. This approach paves the way for agile, retraining-free perception module in intelligent transportation system.

Enterprise environments contain a heterogeneous, rapidly growing collection of internal artifacts related to code, data, and many different tools. Critical information for assessing privacy risk and ensuring regulatory compliance is often embedded across these varied resources, each with their own arcane discovery and extraction techniques. Therefore, large-scale privacy compliance in adherence to governmental regulations requires systems to discern the interconnected nature of diverse artifacts in a common, shared universe. We present Privacy Artifact ConnecT or (PACT), an embeddings-driven graph that links millions of artifacts spanning multiple artifact types generated by a variety of teams and projects. Powered by the state-of-the-art DRAGON embedding model, PACT uses a contrastive learning objective with light fine-tuning to link artifacts via their textual components such as raw metadata, ownership specifics, and compliance context. Experimental results show that PACT's fine-tuned model improves recall@1 from 18% to 53%, the query match rate from 9.6% to 69.7% when paired with a baseline AI agent, and the hitrate@1 from 25.7% to 44.9% for candidate selection in a standard recommender system.

Recent advances in large language models (LLMs) have intensified the need to understand and reliably curb their harmful behaviours. We introduce a multidimensional framework for probing and steering harmful content in model internals. For each of 55 distinct harmfulness subconcepts (e.g., racial hate, employment scams, weapons), we learn a linear probe, yielding 55 interpretable directions in activation space. Collectively, these directions span a harmfulness subspace that we show is strikingly low-rank. We then test ablation of the entire subspace from model internals, as well as steering and ablation in the subspace's dominant direction. We find that dominant direction steering allows for near elimination of harmfulness with a low decrease in utility. Our findings advance the emerging view that concept subspaces provide a scalable lens on LLM behaviour and offer practical tools for the community to audit and harden future generations of language models.

We introduce Inworld TTS-1, a set of two Transformer-based autoregressive text-to-speech (TTS) models. Our largest model, TTS-1-Max, has 8.8B parameters and is designed for utmost quality and expressiveness in demanding applications. TTS-1 is our most efficient model, with 1.6B parameters, built for real-time speech synthesis and on-device use cases. By scaling train-time compute and applying a sequential process of pre-training, fine-tuning, and RL-alignment of the speech-language model (SpeechLM) component, both models achieve state-of-the-art performance on a variety of benchmarks, demonstrating exceptional quality relying purely on in-context learning of the speaker's voice. Inworld TTS-1 and TTS-1-Max can generate high-resolution 48 kHz speech with low latency, and support 11 languages with fine-grained emotional control and non-verbal vocalizations through audio markups. We additionally open-source our training and modeling code under an MIT license.