The rise of large language models (LLMs) has made natural language-driven route planning an emerging research area that encompasses rich user objectives. Current research exhibits two distinct approaches: direct route planning using LLM-as-Agent and graph-based searching strategies. However, LLMs in the former approach struggle to handle extensive map data, while the latter shows limited capability in understanding natural language preferences. Additionally, a more critical challenge arises from the highly heterogeneous and unpredictable spatio-temporal distribution of users across the globe. In this paper, we introduce a novel LLM-Assisted route Planning (LLMAP) system that employs an LLM-as-Parser to comprehend natural language, identify tasks, and extract user preferences and recognize task dependencies, coupled with a Multi-Step Graph construction with iterative Search (MSGS) algorithm as the underlying solver for optimal route finding. Our multi-objective optimization approach adaptively tunes objective weights to maximize points of interest (POI) quality and task completion rate while minimizing route distance, subject to three key constraints: user time limits, POI opening hours, and task dependencies. We conduct extensive experiments using 1,000 routing prompts sampled with varying complexity across 14 countries and 27 cities worldwide. The results demonstrate that our approach achieves superior performance with guarantees across multiple constraints.

This paper proposes the MT-DQN model, which integrates a Transformer, Temporal Graph Neural Network (TGNN), and Deep Q-Network (DQN) to address the challenges of predicting user behavior and optimizing recommendation strategies in short-video environments. Experiments demonstrated that MT-DQN consistently outperforms traditional concatenated models, such as Concat-Modal, achieving an average F1-score improvement of 10.97% and an average NDCG@5 improvement of 8.3%. Compared to the classic reinforcement learning model Vanilla-DQN, MT-DQN reduces MSE by 34.8% and MAE by 26.5%. Nonetheless, we also recognize challenges in deploying MT-DQN in real-world scenarios, such as its computational cost and latency sensitivity during online inference, which will be addressed through future architectural optimization.

Large multimodal models (LMMs) encode universal physical laws observed during training, such as momentum conservation, as parametric knowledge. It allows LMMs to answer physical reasoning queries, such as the outcome of a potential collision event from visual input. However, since parametric knowledge includes only the physical laws seen during training, it is insufficient for reasoning when the inference scenario violates these physical laws. In contrast, humans possess the skill to adapt their physical reasoning to unseen physical environments from a few visual examples. This ability, which we refer to as inductive physical reasoning, is indispensable for LMMs if they are to replace human agents in safety-critical applications. Despite its importance, existing visual benchmarks evaluate only the parametric knowledge in LMMs, and not inductive physical reasoning. To this end, we propose InPhyRe, the first visual question answering benchmark to measure inductive physical reasoning in LMMs. InPhyRe evaluates LMMs on their ability to predict the outcome of collision events in algorithmically generated synthetic collision videos. By inspecting 13 LMMs, InPhyRe informs us that (1) LMMs struggle to apply their limited parametric knowledge about universal physical laws to reasoning, (2) inductive physical reasoning in LMMs is weak when demonstration samples violate universal physical laws, and (3) inductive physical reasoning in LMMs suffers from language bias and largely ignores the visual inputs, questioning the trustworthiness of LMMs regarding visual inputs.

This paper develops an autonomous agentic framework called V-Math that aims to assist Vietnamese high school students in preparing for the National High School Graduation Mathematics Exams (NHSGMEs). The salient framework integrates three specialized AI agents: a specification-matrix-conditioned question generator, a solver/explainer for detailed step-by-step reasoning, and a personalized tutor that adapts to student performance. Beyond enabling self-paced student practice, V-Math supports teachers by generating innovative, compliant exam questions and building diverse, high-quality question banks. This reduces manual workload and enriches instructional resources. We describe the system architecture, focusing on practice modes for learners and teacher-oriented features for question generation. Preliminary evaluations demonstrate that V-Math produces matrix-aligned exams with high solution accuracy, delivers coherent explanations, and enhances the variety of practice materials. These results highlight its potential to support scalable, equitable mathematics preparation aligned with national standards while also empowering teachers through AI-assisted exam creation.

The perception and generation of Human-Object Interaction (HOI) are crucial for fields such as robotics, AR/VR, and human behavior understanding. However, current approaches model this task in an offline setting, where information at each time step can be drawn from the entire interaction sequence. In contrast, in real-world scenarios, the information available at each time step comes only from the current moment and historical data, i.e., an online setting. We find that offline methods perform poorly in an online context. Based on this observation, we propose two new tasks: Online HOI Generation and Perception. To address this task, we introduce the OnlineHOI framework, a network architecture based on the Mamba framework that employs a memory mechanism. By leveraging Mamba's powerful modeling capabilities for streaming data and the Memory mechanism's efficient integration of historical information, we achieve state-of-the-art results on the Core4D and OAKINK2 online generation tasks, as well as the online HOI4D perception task.

The Internet of Electric Vehicles (IoEV) envisions a tightly coupled ecosystem of electric vehicles (EVs), charging infrastructure, and grid services, yet it remains vulnerable to cyberattacks, unreliable battery-state predictions, and opaque decision processes that erode trust and performance. To address these challenges, we introduce a novel Agentic Artificial Intelligence (AAI) framework tailored for IoEV, where specialized agents collaborate to deliver autonomous threat mitigation, robust analytics, and interpretable decision support. Specifically, we design an AAI architecture comprising dedicated agents for cyber-threat detection and response at charging stations, real-time State of Charge (SoC) estimation, and State of Health (SoH) anomaly detection, all coordinated through a shared, explainable reasoning layer; develop interpretable threat-mitigation mechanisms that proactively identify and neutralize attacks on both physical charging points and learning components; propose resilient SoC and SoH models that leverage continuous and adversarial-aware learning to produce accurate, uncertainty-aware forecasts with human-readable explanations; and implement a three-agent pipeline, where each agent uses LLM-driven reasoning and dynamic tool invocation to interpret intent, contextualize tasks, and execute formal optimizations for user-centric assistance. Finally, we validate our framework through comprehensive experiments across diverse IoEV scenarios, demonstrating significant improvements in security and prediction accuracy. All datasets, models, and code will be released publicly.

Fine-tuning large language models (LLMs) with parameter-efficient techniques such as LoRA and QLoRA has enabled adaptation of foundation models on modest hardware. Yet the efficiency of such training on consumer-grade GPUs, especially under strict 8 GB VRAM limits, remains underexplored. We present a controlled profiling study of LoRA/QLoRA fine-tuning using the Qwen2.5-1.5B-Instruct model on a single NVIDIA RTX 4060. Across three representative configurations, we systematically vary batch size, sequence length, optimizer choice (AdamW vs. PagedAdamW), and precision (fp16 vs. bf16). We report throughput (tokens/s), time per 10k tokens, and VRAM footprint, alongside energy estimates derived from GPU board power limits. Our results show that paged optimizers improve throughput by up to 25% (628 tok/s vs. 500 tok/s baseline), while bf16 degrades efficiency relative to fp16. Despite 8 GB constraints, sequence lengths up to 2048 tokens were feasible using parameter-efficient strategies. To our knowledge, this is the first systematic case study of LLM fine-tuning efficiency on consumer GPUs, providing reproducible benchmarks and practical guidelines for resource-constrained researchers and practitioners.

Large language models (LLMs) enforce safety alignment to reliably refuse malicious requests, yet the same blanket safeguards also block legitimate uses in policing, defense, and other high-stakes settings. Earlier "refusal-direction" edits can bypass those layers, but they rely on a single vector that indiscriminately unlocks all hazardous topics, offering no semantic control. We introduce Mutually Exclusive Unlock Vectors (MEUV), a lightweight framework that factorizes the monolithic refusal direction into topic-aligned, nearly orthogonal vectors, each dedicated to one sensitive capability. MEUV is learned in a single epoch with a multi-task objective that blends a differential-ablation margin, cross-topic and orthogonality penalties, and several auxiliary terms. On bilingual malicious-prompt benchmarks, MEUV achieves an attack success rate of no less than 87% on Gemma-2-2B, LLaMA-3-8B, and Qwen-7B, yet cuts cross-topic leakage by up to 90% compared with the best single-direction baseline. Vectors trained in Chinese transfer almost unchanged to English (and vice versa), suggesting a language-agnostic refusal subspace. The results show that fine-grained, topic-level capability activation is achievable with minimal utility loss, paving the way for controlled LLMs deployment in security-sensitive domains.

Serving large language models (LLMs) efficiently remains challenging due to the high memory and latency overhead of key-value (KV) cache access during autoregressive decoding. We present \textbf{TinyServe}, a lightweight and extensible serving system for deploying tiny LLMs (e.g., TinyLLaMA, GPT2-345M) with support for structured KV sparsity, plugin-based token selection, and hardware-efficient attention kernels. Unlike prior simulation frameworks, TinyServe executes real-time decoding with configurable sparsity strategies and fine-grained instrumentation. To reduce decoding cost, we introduce a \textit{query-aware page selection} mechanism that leverages bounding-box metadata to estimate attention relevance between the query and KV cache blocks. This enables selective KV loading with minimal overhead and no model modifications. Our fused CUDA kernel integrates page scoring, sparse memory access, and masked attention in a single pass. Experiments show that TinyServe achieves up to \textbf{3.4x} speedup and over \textbf{2x} memory savings with negligible accuracy drop. Additional analysis of cache reuse, page hit rate, and multi-GPU scaling confirms its practicality as an efficient system-level design for LLM training and inference research on resource-constrained hardware.

In natural language processing tasks, pure reinforcement learning (RL) fine-tuning methods often suffer from inefficient exploration and slow convergence; while supervised fine-tuning (SFT) methods, although efficient in training, have limited performance ceiling and less solid theoretical foundation compared to RL. To address efficiency-capability trade-off, we propose the Guess-Think-Answer (GTA) framework that combines the efficiency of SFT with the capability gains of RL in a unified training paradigm. GTA works by having the model first produce a provisional guess (optimized via cross-entropy loss), then reflect on this guess before generating the final answer, with RL rewards shaping both the final output and the format of the entire GTA structure. This hybrid approach achieves both faster convergence than pure RL and higher performance ceiling than pure SFT. To mitigate gradient conflicts between the two training signals, we employ loss masking and gradient constraints. Empirical results on four text classification benchmarks demonstrate that GTA substantially accelerates convergence while outperforming both standalone SFT and RL baselines.