When users submit queries to Large Language Models (LLMs), their prompts can often contain sensitive data, forcing a difficult choice: Send the query to a powerful proprietary LLM providers to achieving state-of-the-art performance and risk data exposure, or relying on smaller, local models guarantees data privacy but often results in a degradation of task performance. Prior approaches have relied on static pipelines that use LLM rewriting, which shatters linguistic coherence and indiscriminately removes privacy-sensitive information, including task-critical content. We reformulate this challenge (Privacy-Conscious Delegation) as a sequential decision-making problem and introduce a novel reinforcement learning (RL) framework called PrivacyPAD to solve it. Our framework trains an agent to dynamically route text chunks, learning a policy that optimally balances the trade-off between privacy leakage and task performance. It implicitly distinguishes between replaceable Personally Identifiable Information (PII) (which it shields locally) and task-critical PII (which it strategically sends to the remote model for maximal utility). To validate our approach in complex scenarios, we also introduce a new medical dataset with high PII density. Our framework achieves a new state-of-the-art on the privacy-utility frontier, demonstrating the necessity of learned, adaptive policies for deploying LLMs in sensitive environments.

Autonomous agents capable of operating complex graphical user interfaces (GUIs) have the potential to transform desktop automation. While recent advances in large language models (LLMs) have significantly improved UI understanding, navigating full-window, multi-application desktop environments remains a major challenge. Data availability is limited by costly manual annotation, closed-source datasets and surface-level synthetic pipelines. We introduce GUIrilla, an automated scalable framework that systematically explores applications via native accessibility APIs to address the critical data collection challenge in GUI automation. Our framework focuses on macOS - an ecosystem with limited representation in current UI datasets - though many of its components are designed for broader cross-platform applicability. GUIrilla organizes discovered interface elements and crawler actions into hierarchical GUI graphs and employs specialized interaction handlers to achieve comprehensive application coverage. Using the application graphs from GUIrilla crawler, we construct and release GUIrilla-Task, a large-scale dataset of 27,171 functionally grounded tasks across 1,108 macOS applications, each annotated with full-desktop and window-level screenshots, accessibility metadata, and semantic action traces. Empirical results show that tuning LLM-based agents on GUIrilla-Task significantly improves performance on downstream UI tasks, outperforming synthetic baselines on the ScreenSpot Pro benchmark while using 97% less data. We also release macapptree, an open-source library for reproducible collection of structured accessibility metadata, along with the full GUIrilla-Task dataset, the manually verified GUIrilla-Gold benchmark, and the framework code to support open research in desktop autonomy.

Large language models (LLMs) have demonstrated remarkable capabilities in various kinds of tasks, while the billion or even trillion parameters bring storage and efficiency bottlenecks for inference. Quantization, particularly floating-point quantization, is known to be capable of speeding up LLM inference by reducing memory footprint and data movement during the inference process. For the first time, we advance the floating-point quantization exploration from integer bitwidths to non-integer bit-widths, namely AMS-Quant, to further approach the quantization sweet spot. AMS-Quant incorporates two novel techniques to put it into effect: (1) it proposes Mantissa-bit Sharing, which groups k quantized weights and lets them share the least significant mantissa bit, allowing us to further approach the minimum quantization bit-width without accuracy loss. (2) It introduces Adaptive Searching, which employs an offline optimization strategy to minimize the accuracy degradation introduced by sharing. Moreover, AMS-Quant is also prototyped as efficient CUDA Linear kernels, which translates memory savings into wall-clock latency reduction by reducing memory access. Extensive experiments on large-scale datasets and models show that AMS-Quant can quantize the model to FP-5.33-e2m3 and FP4.25-e2m2, and significantly speed up the LLM decoding over FP16 inference (2.8x and 3.2x), with negligible accuracy loss.

Running Large Language Models (LLMs) on edge devices is crucial for reducing latency, improving real-time processing, and enhancing privacy. By performing inference directly on the device, data does not need to be sent to the cloud, ensuring faster responses and reducing reliance on network connectivity. However, implementing LLMs on edge devices presents challenges, particularly with managing key-value (KV) caches, which plays a pivotal role in LLM serving. As the input text lengthens, the size of the KV cache increases linearly with the sequence length, leading to a significant memory footprint and data access costs. On the other hand, edge devices have limited memory and computational power, making it hard to store and efficiently access the large caches needed for LLM inference. To mitigate the substantial overhead caused by KV cache, we propose using embedded DRAM (eDRAM) as the primary storage for LLM serving in edge device, which offers higher storage density compared to SRAM. However, to ensure data integrity, eDRAM needs periodic refresh operations, which are power-intensive. To reduce eDRAM costs and improve overall system performance, we propose~\textit{Kelle}, a software-hardware co-design solution optimized for deploying LLMs on eDRAM-based edge systems. Combined with our fine-grained memory eviction, recomputation, and refresh control algorithms, the \textit{Kelle} accelerator delivers a $3.9\times$ speedup and $4.5\times$ energy savings compared to existing baseline solutions.

Adapting pretrained large language models (LLMs) to code domains via supervised fine-tuning (FT) has been commonly used for code generation. However, we identify a previously underappreciated failure mode, the memorization barrier, where strong memorization of downstream code data in the base model could trap optimization and prevent the standard FT from effectively acquiring new, generalizable code knowledge. To overcome this barrier, we propose the information bottleneck (IB)-guided fine-tuning, termed IB-FT, which applies an IB penalty on hidden representations of the code data to compress spurious, memorized features while preserving task-relevant information. Extensive experiments on two code benchmarks (OriGen and Evol-CodeAlpaca-V1) show that IB-FT substantially alleviates the memorization barrier, improves top-1 performance (Pass@$1$), and yields far more stable gains under the stricter multi-sample metric Pass@$k^{(m)}$ (a problem counts as solved only if at least $m$ of $k$ samples pass unit tests) compared with conventional FT.

Infrastructure in smart cities is increasingly monitored by networks of closed circuit television (CCTV) cameras. Roads, bridges and tunnels develop cracks, potholes, and fluid leaks that threaten public safety and require timely repair. Manual inspection is costly and hazardous, and existing automatic systems typically address individual defect types or provide unstructured outputs that cannot directly guide maintenance crews. This paper proposes a comprehensive pipeline that leverages street CCTV streams for multi defect detection and segmentation using the YOLO family of object detectors and passes the detections to a vision language model (VLM) for scene aware summarization. The VLM generates a structured action plan in JSON format that includes incident descriptions, recommended tools, dimensions, repair plans, and urgent alerts. We review literature on pothole, crack and leak detection, highlight recent advances in large vision language models such as QwenVL and LLaVA, and describe the design of our early prototype. Experimental evaluation on public datasets and captured CCTV clips demonstrates that the system accurately identifies diverse defects and produces coherent summaries. We conclude by discussing challenges and directions for scaling the system to city wide deployments.

AI red teaming brings security thinking to LLM applications by probing failure modes such as prompt injection, output manipulation, and sensitive data exfiltration. While automated and curated benchmarks (e.g., JailbreakBench [1], HarmBench [2]) are increasingly used to test models and defenses, comparatively fewer studies analyze community scale behavior in the wild. We study ai_gon3_rogu3 [3], a 10 day competition with 504 registrants and 217 active players, to quantify solve dynamics, tactic stratification, and choke points across 11 challenges. We find sharp skill stratification, higher success for output manipulation than for data extraction, and strong effects of format obfuscation tactics, with multi step defenses remaining robust, among other insights.

Most financial recommendation systems often fail to account for key behavioral and regulatory factors, leading to advice that is misaligned with user preferences, difficult to interpret, or unlikely to be followed. We present FLARKO (Financial Language-model for Asset Recommendation with Knowledge-graph Optimization), a novel framework that integrates Large Language Models (LLMs), Knowledge Graphs (KGs), and Kahneman-Tversky Optimization (KTO) to generate asset recommendations that are both profitable and behaviorally aligned. FLARKO encodes users' transaction histories and asset trends as structured KGs, providing interpretable and controllable context for the LLM. To demonstrate the adaptability of our approach, we develop and evaluate both a centralized architecture (CenFLARKO) and a federated variant (FedFLARKO). To our knowledge, this is the first demonstration of combining KTO for fine-tuning of LLMs for financial asset recommendation. We also present the first use of structured KGs to ground LLM reasoning over behavioral financial data in a federated learning (FL) setting. Evaluated on the FAR-Trans dataset, FLARKO consistently outperforms state-of-the-art recommendation baselines on behavioral alignment and joint profitability, while remaining interpretable and resource-efficient.

The growing industrial demand for customized and cost-efficient large language models (LLMs) is fueled by the rise of vertical, domain-specific tasks and the need to optimize performance under constraints such as latency and budget. Knowledge distillation, as an efficient model compression and transfer technique, offers a feasible solution. However, existing distillation frameworks often require manual intervention and struggle to meet such complex user-defined distillation requirements. To bridge this gap, we propose Stratos, an end-to-end LLM distillation pipeline that automates server and model selection, knowledge distillation, and deployment in distributed cloud environments. Given user-defined constraints on model performance and system budget, Stratos automatically selects Pareto-optimal servers, dynamically matches teacher-student pairs, and adapts distillation strategies based on task complexity to optimize cloud hosting. Experiments show that Stratos produces a student model that achieves four times the accuracy of its GPT-4o teacher baseline on a rare, domain-specific Mahjong reasoning task with reverse synthetic data and knowledge injection. Moreover, it achieves reduced latency and cost without compromising accuracy. These results highlight its promise for vertical-domain LLM deployment.

Reinforcement Learning with Verifiable Rewards (RLVR), primarily driven by the Group Relative Policy Optimization (GRPO) algorithm, is a leading approach for enhancing the reasoning abilities of Large Language Models (LLMs). Despite its wide adoption, GRPO's gains are often inconsistent; for instance, a model may show significant improvement in one reasoning domain, like mathematics, yet remain stagnant in another, such as medicine. This inconsistency raises a critical question: under what conditions does GRPO improve reasoning and generalize out-of-distribution (OOD)? We investigate this from a data distribution perspective. We first prove theoretically that GRPO is a conservative reweighting scheme, bounded by the base model's distribution and thus unable to discover completely novel solutions. We further validate this in carefully designed controlled studies by training transformers from scratch, evaluating generalization across reasoning depth, input length, token representation, and compositionality. Our results provide a principled explanation for GRPO's boundaries: OOD improvement emerges only when the target task aligns with the model's pretrained biases, while gains on in-distribution (ID) tasks diminish as performance saturates. This reframes GRPO not as a universal reasoning enhancer but as a tool that sharpens pretraining biases. Our findings motivate future development of algorithms that can expand a model's capabilities beyond its pretraining origin.