Large language models (LLMs) are increasingly deployed as agents in dynamic, real-world environments, where success requires both reasoning and effective tool use. A central challenge for agentic tasks is the growing context length, as agents must accumulate long histories of actions and observations. This expansion raises costs and reduces efficiency in long-horizon tasks, yet prior work on context compression has mostly focused on single-step tasks or narrow applications. We introduce Agent Context Optimization (ACON), a unified framework that optimally compresses both environment observations and interaction histories into concise yet informative condensations. ACON leverages compression guideline optimization in natural language space: given paired trajectories where full context succeeds but compressed context fails, capable LLMs analyze the causes of failure, and the compression guideline is updated accordingly. Furthermore, we propose distilling the optimized LLM compressor into smaller models to reduce the overhead of the additional module. Experiments on AppWorld, OfficeBench, and Multi-objective QA show that ACON reduces memory usage by 26-54% (peak tokens) while largely preserving task performance, preserves over 95% of accuracy when distilled into smaller compressors, and enhances smaller LMs as long-horizon agents with up to 46% performance improvement. Our code is available at https://github.com/microsoft/acon.

As LLM-based agents are increasingly deployed in real-life scenarios, existing benchmarks fail to capture their inherent complexity of handling extensive information, leveraging diverse resources, and managing dynamic user interactions. To address this gap, we introduce VitaBench, a challenging benchmark that evaluates agents on versatile interactive tasks grounded in real-world settings. Drawing from daily applications in food delivery, in-store consumption, and online travel services, VitaBench presents agents with the most complex life-serving simulation environment to date, comprising 66 tools. Through a framework that eliminates domain-specific policies, we enable flexible composition of these scenarios and tools, yielding 100 cross-scenario tasks (main results) and 300 single-scenario tasks. Each task is derived from multiple real user requests and requires agents to reason across temporal and spatial dimensions, utilize complex tool sets, proactively clarify ambiguous instructions, and track shifting user intent throughout multi-turn conversations. Moreover, we propose a rubric-based sliding window evaluator, enabling robust assessment of diverse solution pathways in complex environments and stochastic interactions. Our comprehensive evaluation reveals that even the most advanced models achieve only 30% success rate on cross-scenario tasks, and less than 50% success rate on others. Overall, we believe VitaBench will serve as a valuable resource for advancing the development of AI agents in practical real-world applications. The code, dataset, and leaderboard are available at https://vitabench.github.io/

Social robots play a key role in many applications such as elderly care, home assistant, customer service, and education where they assist, interact, and communicate with humans in a socially intelligent manner. These robots must ensure not only physical safety but also psychological comfort for humans by following social norms. For instance, a robot should avoid disrupting a group conversation when navigating a crowded space as this could be seen as impolite or intrusive. To accomplish this, the robot must not only detect humans but also recognize and interpret their interactions such as conversations, discussions, gatherings, and collaborative activities to adapt its movements accordingly. According to [1, 2], human group interactions are structured into three distinct spaces: (i) o-space, the central region where active participants focus their attention, (ii) p-space, the surrounding area occupied by engaged individuals, and (iii) r-space, the outer region where bystanders or non-participants are positioned. To enable socially aware navigation, recognition algorithms must estimate these spatial regions.

Quantization techniques are essential for the deployment of Large Language Models (LLMs) on edge devices. However, prevailing methods often rely on mixed-precision multiplication that lacks efficient hardware support, making it not feasible. However, such aggressive compression leads to significant accuracy degradation, even after costly quantization-aware training with massive data. We identify the core issue as deadzone trapping: a large number of weights are trapped at the dead-zone boundary. This occurs because these weights receive only noisy, uninformative gradients, preventing stable escape from the deadzone and severely impeding model capacity and optimization. To address this issue, we propose T equila, a trapping-free quantization optimization method that reactivates deadzone-trapped weights by repurposing them as dynamic biases. This allows the repurposed weights to provide a continuous signal in the forward pass and, critically, receive direct, meaningful gradient signals during backpropagation, thereby enhancing model capacity and optimization with nearly zero inference overhead. Extensive evaluations demonstrate that Tequila outperforms state-of-the-art (SOT A) ternary quantization methods across five benchmarks. Specifically, on the ARC benchmark, it achieves > 4% accuracy gain over the SOT A baseline, nearly matching full-precision performance (within < 1% gap) with a 3.0 inference speedup. Consequently, Tequila offers a highly practical and efficient implementation for the deployment of advanced LLMs in resource-constrained environments. Recent advancements in large language models (LLMs) (Wu et al., 2023; Floridi & Chiriatti, 2020; Zhang et al., 2022) have demonstrated remarkable success across a wide range of applications, from conversational chatbots to creative writing.

Multimodal large language models (MLLMs) have recently achieved remarkable progress in radiology by integrating visual perception with natural language understanding. However, they often generate clinically unsupported descriptions, known as medical hallucinations, which pose serious risks in medical applications that demand accuracy and image-grounded outputs. Through empirical analysis, we find that prompt-induced hallucinations remain prevalent in radiology MLLMs, largely due to over-sensitivity to clinical sections. To address this, we introduce Clinical Contrastive Decoding (CCD), a training-free and retrieval-free inference framework that integrates structured clinical signals from task-specific radiology expert models. CCD introduces a dual-stage contrastive mechanism to refine token-level logits during generation, thereby enhancing clinical fidelity without modifying the base MLLM. Experiments on three datasets and multiple models demonstrate that CCD consistently improves overall performance on radiology report generation (RRG). On the MIMIC-CXR dataset, it yields up to a 17% improvement in RadGraph-F1 when applied to state-of-the-art RRG models. Our approach provides a lightweight and generalisable solution for mitigating medical hallucinations, effectively bridging expert models and MLLMs in radiology.

Parameter-efficient fine-tuning (PEFT) of powerful pre-trained models for complex downstream tasks has proven effective in vision and language processing, yet this paradigm remains unexplored in scientific machine learning, where the objective is to model complex physical systems. We conduct the first systematic study of PEFT for pre-trained Large Operator Models (LOMs) obtained by scaling variants of Fourier Neural Operator. First, we observe that the widely used Low-Rank Adaptation (LoRA) yields markedly poorer performance on LOMs than Adapter tuning. Then, we further theoretically establish that stacked LoRA incurs a depth-amplified lower bound on approximation error within Fourier layers, whereas adapters retain universal approximation capacity and, by concentrating parameters on energy-dominant low-frequency modes, attain exponentially decaying error with bottleneck width in the Fourier domain. Motivated by the robust empirical gains of adapters and by our theoretical characterization of PDE solutions as spectrally sparse, we introduce Frequency-Adaptive Adapter (F-Adapter). F-Adapter allocates adapter capacity based on spectral complexity, assigning higher-dimension modules to low-frequency components and lower-dimension modules to high-frequency components. Our F-Adapters establish state-of-the-art (SOTA) results on multiple challenging 3D Navier-Stokes benchmarks, markedly enhancing both generalization and spectral fidelity over LoRA and other PEFT techniques commonly used in LLMs. To the best of our knowledge, this work is the first to explore PEFT for scientific machine-learning and establishes F-Adapter as an effective paradigm for this domain.

Abstract--In mixed-traffic environments, where autonomous vehicles (A Vs) must interact with diverse human-driven vehicles (HVs), the unpredictability of human intentions and heterogeneous driving behaviors poses significant challenges to safe and efficient lane change maneuvers. Existing methods often oversimplify these interactions by assuming uniform or fixed behavioral patterns. T o address this limitation, we propose an intention-driven lane change framework that integrates driving-style recognition with cooperation-aware decision-making and motion-planning. First, a deep learning-based classifier is developed to identify distinct human driving styles from the NGSIM dataset in real time. Second, we introduce a cooperation score composed of intrinsic and interactive components, which estimates surrounding drivers' intentions and quantifies their willingness to cooperate with the ego vehicle's lane change. Third, a decision-making module is designed by combining behavior cloning (BC) with inverse reinforcement learning (IRL) to determine whether a lane change should be initiated under current conditions. Finally, a coordinated motion-planning architecture is established, integrating IRL-based intention inference with model predictive control (MPC) to generate collision-free and socially compliant trajectories. Extensive experiments demonstrate that the proposed intention-driven BC-IRL model achieves superior performance, reaching 94.2% accuracy and 94.3% F1-score, and outperforming multiple rule-based and learning-based baselines. In particular, it improves lane change recognition by 4-15% in F1-score, highlighting the benefit of modeling inter-driver heterogeneity via intrinsic and interactive cooperation scores.

The immense computational cost of training Large Language Models (LLMs) presents a major barrier to innovation. While FP8 training offers a promising solution with significant theoretical efficiency gains, its widespread adoption has been hindered by the lack of a comprehensive, open-source training recipe. To bridge this gap, we introduce an end-to-end FP8 training recipe that seamlessly integrates continual pre-training and supervised fine-tuning. Our methodology employs a fine-grained, hybrid-granularity quantization strategy to maintain numerical fidelity while maximizing computational efficiency. Through extensive experiments, including the continue pre-training of models on a 160B-token corpus, we demonstrate that our recipe is not only remarkably stable but also essentially lossless, achieving performance on par with the BF16 baseline across a suite of reasoning benchmarks. Crucially, this is achieved with substantial efficiency improvements, including up to a 22% reduction in training time, a 14% decrease in peak memory usage, and a 19% increase in throughput. Our results establish FP8 as a practical and robust alternative to BF16, and we will release the accompanying code to further democratize large-scale model training.

The widespread adoption of Large Language Models (LLMs) has revolutionized AI deployment, enabling autonomous and semi-autonomous applications across industries through intuitive language interfaces and continuous improvements in model development. However, the attendant increase in autonomy and expansion of access permissions among AI applications also make these systems compelling targets for malicious attacks. Their inherent susceptibility to security flaws necessitates robust defenses, yet no known approaches can prevent zero-day or novel attacks against LLMs. This places AI protection systems in a category similar to established malware protection systems: rather than providing guaranteed immunity, they minimize risk through enhanced observability, multi-layered defense, and rapid threat response, supported by a threat intelligence function designed specifically for AI-related threats. Prior work on LLM protection has largely evaluated individual detection models rather than end-to-end systems designed for continuous, rapid adaptation to a changing threat landscape. We present a production-grade defense system rooted in established malware detection and threat intelligence practices. Our platform integrates three components: a threat intelligence system that turns emerging threats into protections; a data platform that aggregates and enriches information while providing observability, monitoring, and ML operations; and a release platform enabling safe, rapid detection updates without disrupting customer workflows. Together, these components deliver layered protection against evolving LLM threats while generating training data for continuous model improvement and deploying updates without interrupting production.

This paper introduces a simple and scalable approach to improve the data efficiency of large language model (LLM) training by augmenting existing text data with thinking trajectories. The compute for pre-training LLMs has been growing at an unprecedented rate, while the availability of high-quality data remains limited. Consequently, maximizing the utility of available data constitutes a significant research challenge. A primary impediment is that certain high-quality tokens are difficult to learn given a fixed model capacity, as the underlying rationale for a single token can be exceptionally complex and deep. To address this issue, we propose Thinking augmented Pre-Training (TPT), a universal methodology that augments text with automatically generated thinking trajectories. Such augmentation effectively increases the volume of the training data and makes high-quality tokens more learnable through step-by-step reasoning and decomposition. We apply TPT across diverse training configurations up to $100$B tokens, encompassing pre-training with both constrained and abundant data, as well as mid-training from strong open-source checkpoints. Experimental results indicate that our method substantially improves the performance of LLMs across various model sizes and families. Notably, TPT enhances the data efficiency of LLM pre-training by a factor of $3$. For a $3$B parameter model, it improves the post-training performance by over $10\%$ on several challenging reasoning benchmarks.