--The integration of wireless communications and Large Language Models (LLMs) is poised to unlock ubiquitous intelligent services, yet deploying them in wireless edge-device collaborative environments presents a critical trade-off between inference quality and end-to-end latency. A fundamental mismatch exists between task complexity and resource allocation: offloading simple queries invites prohibitive latency, while on-device models lack the capacity for demanding computations. T o address this challenge, we propose a dynamic, quality-latency aware routing framework that orchestrates inference between a lightweight model on the mobile device and a powerful model on the edge server . Our framework employs two distinct cost models: for single-turn queries, it fuses a BERT -predicted semantic score with communication and computation overheads; for multi-turn dialogues, it further quantifies context-aware costs arising from model switching and KV-cache management. While maintaining full inference quality, extensive experiments demonstrate that our framework cuts average response latency by 5-15% and reduces large model invocations by 10-20% against competitive baselines on MMLU, GSM8K, and MT -Bench-101 benchmarks.

LLMs increasingly exhibit over-refusal behavior, where safety mechanisms cause models to reject benign instructions that superficially resemble harmful content. This phenomena diminishes utility in production applications that repeatedly rely on common prompt templates or applications that frequently rely on LLMs for specific tasks (e.g. sentiment analysis, language translation). Through comprehensive evaluation, we demonstrate that LLMs still tend to refuse responses to harmful instructions when those instructions are reframed to appear as benign tasks. Our mechanistic analysis reveal that LLMs follow distinct "constellation" patterns in embedding space as representations traverse layers, with each task maintaining consistent trajectories that shift predictably between refusal and non-refusal cases. We introduce SafeConstellations, an inference-time trajectory-shifting approach that tracks task-specific trajectory patterns and guides representations toward non-refusal pathways. By selectively guiding model behavior only on tasks prone to over-refusal, and by preserving general model behavior, our method reduces over-refusal rates by up to 73% with minimal impact on utility-offering a principled approach to mitigating over-refusals.

While deploying large language models on edge devices promises low-latency and privacy-preserving AI services, it is hindered by limited device resources. Although pipeline parallelism facilitates distributed inference, existing approaches often ignore the cold-start latency caused by on-demand model loading. In this paper, we propose a latency-aware scheduling framework that overlaps model loading with computation and communication to minimize total inference latency. Based on device and model parameters, the framework dynamically adjusts layer partitioning and allocation to effectively hide loading time, thereby eliminating as many idle periods as possible. We formulate the problem as a Mixed-Integer Non-Linear Program and design an efficient dynamic programming algorithm to optimize model partitioning and device assignment. Experimental results show that the proposed method significantly reduces cold-start latency compared to baseline strategies.

When humans perform everyday tasks, we naturally adjust our actions based on the current state of the environment. For instance, if we intend to put something into a drawer but notice it is closed, we open it first. However, many autonomous robots lack this adaptive awareness. They often follow pre-planned actions that may overlook subtle yet critical changes in the scene, which can result in actions being executed under outdated assumptions and eventual failure. While replanning is critical for robust autonomy, most existing methods respond only after failures occur, when recovery may be inefficient or infeasible. While proactive replanning holds promise for preventing failures in advance, current solutions often rely on manually designed rules and extensive supervision. In this work, we present a proactive replanning framework that detects and corrects failures at subtask boundaries by comparing scene graphs constructed from current RGB-D observations against reference graphs extracted from successful demonstrations. When the current scene fails to align with reference trajectories, a lightweight reasoning module is activated to diagnose the mismatch and adjust the plan. Experiments in the AI2-THOR simulator demonstrate that our approach detects semantic and spatial mismatches before execution failures occur, significantly improving task success and robustness.

Depression, anxiety, and stress are widespread mental health concerns that increasingly drive individuals to seek information from Large Language Models (LLMs). This study investigates how eight LLMs (Claude Sonnet, Copilot, Gemini Pro, GPT-4o, GPT-4o mini, Llama, Mixtral, and Perplexity) reply to twenty pragmatic questions about depression, anxiety, and stress when those questions are framed for six user profiles (baseline, woman, man, young, old, and university student). The models generated 2,880 answers, which we scored for sentiment and emotions using state-of-the-art tools. Our analysis revealed that optimism, fear, and sadness dominated the emotional landscape across all outputs, with neutral sentiment maintaining consistently high values. Gratitude, joy, and trust appeared at moderate levels, while emotions such as anger, disgust, and love were rarely expressed. The choice of LLM significantly influenced emotional expression patterns. Mixtral exhibited the highest levels of negative emotions including disapproval, annoyance, and sadness, while Llama demonstrated the most optimistic and joyful responses. The type of mental health condition dramatically shaped emotional responses: anxiety prompts elicited extraordinarily high fear scores (0.974), depression prompts generated elevated sadness (0.686) and the highest negative sentiment, while stress-related queries produced the most optimistic responses (0.755) with elevated joy and trust. In contrast, demographic framing of queries produced only marginal variations in emotional tone. Statistical analyses confirmed significant model-specific and condition-specific differences, while demographic influences remained minimal. These findings highlight the critical importance of model selection in mental health applications, as each LLM exhibits a distinct emotional signature that could significantly impact user experience and outcomes.

Vision-language models (VLM) align images and text in a shared representation space that is useful for retrieval and zero-shot transfer . Y et, this alignment can encode and amplify social stereotypes in subtle ways that are not obvious from standard accuracy metrics. In this study, we test whether the contrastive vision-language encoder exhibits gender-linked associations when it places embeddings of face images near embeddings of short phrases that describe occupations and activities. W e assemble a dataset of 220 face photographs split by perceived binary gender and a set of 150 unique statements distributed across six categories covering emotional labor, cognitive labor, domestic labor, technical labor, professional roles, and physical labor . W e compute unit-norm image embeddings for every face and unit-norm text embeddings for every statement, then define a statement-level association score as the difference between the mean cosine similarity to the male set and the mean cosine similarity to the female set, where positive values indicate stronger association with the male set and negative values indicate stronger association with the female set. W e attach bootstrap confidence intervals by re-sampling images within each gender group, aggregate by category with a separate bootstrap over statements, and run a label-swap null model that estimates the level of mean absolute association we would expect if no gender structure were present. The outcome is a statement-wise and category-wise map of gender associations in a contrastive vision-language space, accompanied by uncertainty, simple sanity checks, and a robust gender bias evaluation framework.

Large language models (LLMs) have demonstrated potential in reasoning tasks, but their performance on linguistics puzzles remains consistently poor. These puzzles, often derived from Linguistics Olympiad (LO) contests, provide a minimal contamination environment to assess LLMs' linguistic reasoning abilities across low-resource languages. This work analyses LLMs' performance on 629 problems across 41 low-resource languages by labelling each with linguistically informed features to unveil weaknesses. Our analyses show that LLMs struggle with puzzles involving higher morphological complexity and perform better on puzzles involving linguistic features that are also found in English. We also show that splitting words into morphemes as a pre-processing step improves solvability, indicating a need for more informed and language-specific tokenisers. These findings thus offer insights into some challenges in linguistic reasoning and modelling of low-resource languages.

Instruction fine-tuned large language models (LLMs) enable a simple zero-shot or few-shot prompting paradigm, also known as in-context learning, for building prediction models. This convenience, combined with continued advances in LLM capability, has the potential to drive their adoption across a broad range of domains, including high-stakes applications where group fairness -- preventing disparate impacts across demographic groups -- is essential. The majority of existing approaches to enforcing group fairness on LLM-based classifiers rely on traditional fair algorithms applied via model fine-tuning or head-tuning on final-layer embeddings, but they are no longer applicable to closed-weight LLMs under the in-context learning setting, which include some of the most capable commercial models today, such as GPT-4, Gemini, and Claude. In this paper, we propose a framework for deriving fair classifiers from closed-weight LLMs via prompting: the LLM is treated as a feature extractor, and features are elicited from its probabilistic predictions (e.g., token log probabilities) using prompts strategically designed for the specified fairness criterion to obtain sufficient statistics for fair classification; a fair algorithm is then applied to these features to train a lightweight fair classifier in a post-hoc manner. Experiments on five datasets, including three tabular ones, demonstrate strong accuracy-fairness tradeoffs for the classifiers derived by our framework from both open-weight and closed-weight LLMs; in particular, our framework is data-efficient and outperforms fair classifiers trained on LLM embeddings (i.e., head-tuning) or from scratch on raw tabular features.

Large Language Models (LLMs) have shown significant potential in automating code generation tasks offering new opportunities across software engineering domains. However, their practical application remains limited due to hallucinations - outputs that appear plausible but are factually incorrect, unverifiable or nonsensical. This paper investigates hallucination phenomena in the context of code generation with a specific focus on the automotive domain. A case study is presented that evaluates multiple code LLMs for three different prompting complexities ranging from a minimal one-liner prompt to a prompt with Covesa Vehicle Signal Specifications (VSS) as additional context and finally to a prompt with an additional code skeleton. The evaluation reveals a high frequency of syntax violations, invalid reference errors and API knowledge conflicts in state-of-the-art models GPT-4.1, Codex and GPT-4o. Among the evaluated models, only GPT-4.1 and GPT-4o were able to produce a correct solution when given the most context-rich prompt. Simpler prompting strategies failed to yield a working result, even after multiple refinement iterations. These findings highlight the need for effective mitigation techniques to ensure the safe and reliable use of LLM generated code, especially in safety-critical domains such as automotive software systems.

Large Reasoning Models (LRMs) have demonstrated remarkable problem-solving abilities in mathematics, as evaluated by existing benchmarks exclusively on well-defined problems. However, such evaluation setup constitutes a critical gap, since a genuine intelligent agent should not only solve problems (as a math quiz solver), but also be able~to ask for information when the problems lack sufficient information, enabling proactivity in responding users' requests. To bridge such gap, we proposes a new dataset consisting of two types of incomplete problems with diverse contexts. Based on the dataset, our systematical evaluation of LRMs reveals their inability in proactively asking for information. In addition, we uncover the behaviors related to overthinking and hallucination of LRMs, and highlight the potential and challenges of supervised fine-tuning in learning such ability. We hope to provide new insights in developing LRMs with genuine intelligence, rather than just solving problems.