Consistency in the output of language models is critical for their reliability and practical utility. Due to their training objective, language models learn to model the full space of possible continuations, leading to outputs that can vary significantly in style and content, even for similar or repeated inputs. To address this, we propose a novel decoding algorithm that enhances response consistency across different prompts with no degradation in response quality. By incorporating a latent variable into the next-token sampling process based on the Gumbel reparametrisation trick, our method outperforms standard sampling by up to 10% across semantic and stylistic consistency benchmarks.

Next-generation instruments for ground-based gamma-ray astronomy are marked by a substantial increase in complexity, featuring dozens of telescopes. This leap in scale introduces significant challenges in managing system operations and offline data analysis. Methods, which depend on advanced personnel training and sophisticated software, become increasingly strained as system complexity grows, making it more challenging to effectively support users in such a multifaceted environment. To address these challenges, we propose the development of AI agents based on instruction-finetuned large language models (LLMs). These agents align with specific documentation and codebases, understand the environmental context, operate with external APIs, and communicate with humans in natural language. Leveraging the advanced capabilities of modern LLMs, which can process and retain vast amounts of information, these AI agents offer a transformative approach to system management and data analysis by automating complex tasks and providing intelligent assistance. We present two prototypes that integrate with the Cherenkov Telescope Array Observatory pipelines for operations and offline data analysis. The first prototype automates data model implementation and maintenance for the Configuration Database of the Array Control and Data Acquisition (ACADA). The second prototype is an open-access code generation application tailored for data analysis based on the Gammapy framework.

This paper poses two critical issues in evaluating base models (without post-training): (1) Unstable evaluation during training: in the early stages of pre-training, the models lack the capability to answer questions as required, leading to unstable evaluation results. This instability makes it difficult to provide solid conclusions to guide the training, especially for key experiments such as data ablation and scaling law. (2) Inconsistency between base and instruct models: base models generally exhibit poorer evaluation performance compared to corresponding instruct models. This gap poses a challenge for assessing whether a base model with better evaluation can truly lead to a better instruct model. To address these issues, we propose Base model Oriented Systematic Evaluation (BOSE), a method specifically designed to optimize the evaluation of base models. Specifically, BOSE introduces two key innovations: In-Context Light-instruction Prompt (ICLiP) for open-ended tasks and Blank-ppl for multi-choice tasks with candidate options, which transforms the standard perplexity (ppl) metric into a fill-in-the-blank format to mitigate early-stage evaluation fluctuations. Furthermore, we are the first to propose Kendall's rank correlation to quantitatively measure the evaluation stability and consistency. Experimental results demonstrate that BOSE significantly enhances both the stability of evaluations during pre-training and the consistency between base and instruct models, thereby providing more reliable guidance for the LLMs' training.

Fuzz testing is a crucial component of software security assessment, yet its effectiveness heavily relies on valid fuzz drivers and diverse seed inputs. Recent advancements in Large Language Models (LLMs) offer transformative potential for automating fuzz testing (LLM4Fuzz), particularly in generating drivers and seeds. However, current LLM4Fuzz solutions face critical reliability challenges, including low driver validity rates and seed quality trade-offs, hindering their practical adoption. This paper aims to examine the reliability bottlenecks of LLM-driven fuzzing and explores potential research directions to address these limitations. It begins with an overview of the current development of LLM4SE and emphasizes the necessity for developing reliable LLM4Fuzz solutions. Following this, the paper envisions a vision where reliable LLM4Fuzz transforms the landscape of software testing and security for industry, software development practitioners, and economic accessibility. It then outlines a road ahead for future research, identifying key challenges and offering specific suggestions for the researchers to consider. This work strives to spark innovation in the field, positioning reliable LLM4Fuzz as a fundamental component of modern software testing.

Our method enables training robot policies without collecting any robot data. We first collect human video demonstrations in diverse environments and use inpainting to remove the human hand. A rendered robot is then inserted into the scene using the estimated hand pose. The resulting augmented dataset is used to train an imitation learning policy, which is deployed zero-shot on a real robot. Abstract --Scaling robotics data collection is critical to advancing general-purpose robots. Current approaches often rely on teleoperated demonstrations which are difficult to scale. We propose a novel data collection method that eliminates the need for robotics hardware by leveraging human video demonstrations. By training imitation learning policies on this human data, our approach enables zero-shot deployment on robots without collecting any robot-specific data. T o bridge the embodiment gap between human and robot appearances, we utilize a data editing approach on the input observations that aligns the image distributions between training data on humans and test data on robots. Our method significantly reduces the cost of diverse data collection by allowing anyone with an RGBD camera to contribute. We demonstrate that our approach works in diverse, unseen environments and on varied tasks. I NTRODUCTION Data scarcity remains a key challenge in advancing robotics research. While large-scale data collection efforts are gaining momentum, even the largest robotics datasets [1, 7] are significantly smaller than those used to train generalist models in natural language processing and computer vision. These efforts are constrained by the slow and costly process of collecting data with robotics hardware.

Personalized tool utilization is essential for aligning large language models (LLMs) with user preference in interaction scenarios with various tools. However, most of the current benchmarks primarily focus on either personalization of text generation or direct tool-utilizing, without considering both. In this work, we introduce a novel benchmark ETAPP for evaluating personalized tool invocation, establishing a sandbox environment, and a comprehensive dataset of 800 testing cases covering diverse user profiles. To improve the accuracy of our evaluation, we propose a key-point-based LLM evaluation method, mitigating biases in the LLM-as-a-judge system by manually annotating key points for each test case and providing them to LLM as the reference. Additionally, we evaluate the excellent LLMs and provide an in-depth analysis. Furthermore, we investigate the impact of different tool-invoking strategies on LLMs' personalization performance and the effects of fine-tuning in our task. The effectiveness of our preference-setting and key-point-based evaluation method is also validated. Our findings offer insights into improving personalized LLM agents. Our Code is available at https://github.com/hypasd-art/ETAPP.

We are witnessing a new era where problem-solving and cognitive tasks are being increasingly delegated to Large Language Models (LLMs) across diverse domains, ranging from code generation to holiday planning. This trend also creates a demand for the ubiquitous execution of LLM-powered applications in a wide variety of environments in which traditional terrestrial 2D networking infrastructures may prove insufficient. A promising solution in this context is to extend edge computing into a 3D setting to include aerial platforms organized in multiple layers, a paradigm we refer to as air computing, to augment local devices for running LLM and Generative AI (GenAI) applications. This approach alleviates the strain on existing infrastructure while enhancing service efficiency by offloading computational tasks to the corresponding air units such as UAVs. Furthermore, the coordinated deployment of various air units can significantly improve the Quality of Experience (QoE) by ensuring seamless, adaptive, and resilient task execution. In this study, we investigate the synergy between LLM-based applications and air computing, exploring their potential across various use cases. Additionally, we present a disaster response case study demonstrating how the collaborative utilization of LLMs and air computing can significantly improve outcomes in critical situations.

Predicting the near-term behavior of a reactive agent is crucial in many robotic scenarios, yet remains challenging when observations of that agent are sparse or intermittent. Vision-Language Models (VLMs) offer a promising avenue by integrating textual domain knowledge with visual cues, but their one-shot predictions often miss important edge cases and unusual maneuvers. Our key insight is that iterative, counterfactual exploration--where a dedicated module probes each proposed behavior hypothesis, explicitly represented as a plausible trajectory, for overlooked possibilities--can significantly enhance VLM-based behavioral forecasting. We present TRACE (Tree-of-thought Reasoning And Counterfactual Exploration), an inference framework that couples tree-of-thought generation with domain-aware feedback to refine behavior hypotheses over multiple rounds. Concretely, a VLM first proposes candidate trajectories for the agent; a counterfactual critic then suggests edge-case variations consistent with partial observations, prompting the VLM to expand or adjust its hypotheses in the next iteration. This creates a self-improving cycle where the VLM progressively internalizes edge cases from previous rounds, systematically uncovering not only typical behaviors but also rare or borderline maneuvers, ultimately yielding more robust trajectory predictions from minimal sensor data. We validate TRACE on both ground-vehicle simulations and real-world marine autonomous surface vehicles. Experimental results show that our method consistently outperforms standard VLM-driven and purely model-based baselines, capturing a broader range of feasible agent behaviors despite sparse sensing. Evaluation videos and code are available at trace-robotics.github.io.

Light decoder-based solvers have gained popularity for solving vehicle routing problems (VRPs) due to their efficiency and ease of integration with reinforcement learning algorithms. This paper revisits light decoder-based approaches, analyzing the implications of their reliance on static embeddings and the inherent challenges that arise. Specifically, we demonstrate that in the light decoder paradigm, the encoder is implicitly tasked with capturing information for all potential decision scenarios during solution construction within a single set of embeddings, resulting in high information density. Furthermore, our empirical analysis reveals that the overly simplistic decoder struggles to effectively utilize this dense information, particularly as task complexity increases, which limits generalization to out-of-distribution (OOD) settings. Building on these insights, we show that enhancing the decoder capacity, with a simple addition of identity mapping and a feed-forward layer, can considerably alleviate the generalization issue. Experimentally, our method significantly enhances the OOD generalization of light decoder-based approaches on large-scale instances and complex VRP variants, narrowing the gap with the heavy decoder paradigm. Our code is available at: https://github.com/ V ehicle Routing Problems (VRPs) are a fundamental class of NP-hard combinatorial optimization problems (COPs) with wide-ranging applications in logistics (Konstantakopoulos et al., 2022), transportation (Garaix et al., 2010), and supply chain management (Dondo et al., 2011). Efficiently solving VRPs is critical for reducing operational costs and enhancing service quality in practice. Traditionally, VRPs have been tackled either using exact solvers (e.g., Gurobi) or heuristic solvers (e.g., LKH-3 (Helsgaun, 2017)). While these methods can yield high-quality (or even optimal) solutions for small to moderate-sized instances, they often face challenges in scaling to larger problem sizes or adapting to different problem variants without extensive domain expertise or manual tuning. Neural solvers have emerged as a promising alternative by leveraging advanced deep learning techniques to learn solution strategies directly from data (Bengio et al., 2021). Numerous neural solvers have been proposed for solving VRPs (Bogyrbayeva et al., 2024), with autoregressive construction solvers gaining particular popularity. These solvers sequentially build solutions by adding one feasible node at a time and are valued for their conceptual simplicity and flexibility across different VRP variants. Among them, (heavy encoder) light decoder-based solvers (Vinyals et al., 2015; Kool et al., 2019; Kwon et al., 2020; Kim et al., 2022; Gao et al., 2024; Liu et al., 2024) stand out for their computational efficiency and ease of integration with reinforcement learning (RL) algorithms.

Generating knowledge-intensive and comprehensive long texts, such as encyclopedia articles, remains significant challenges for Large Language Models. It requires not only the precise integration of facts but also the maintenance of thematic coherence throughout the article. Existing methods, such as direct generation and multi-agent discussion, often struggle with issues like hallucinations, topic incoherence, and significant latency. To address these challenges, we propose RAPID, an efficient retrieval-augmented long text generation framework. RAPID consists of three main modules: (1) Retrieval-augmented preliminary outline generation to reduce hallucinations, (2) Attribute-constrained search for efficient information discovery, (3) Plan-guided article generation for enhanced coherence. Extensive experiments on our newly compiled benchmark dataset, FreshWiki-2024, demonstrate that RAPID significantly outperforms state-of-the-art methods across a wide range of evaluation metrics (e.g. long-text generation, outline quality, latency, etc). Our work provides a robust and efficient solution to the challenges of automated long-text generation.