id int64 1 361k | content stringlengths 1 23.3M ⌀ | title stringlengths 1 271 | appendix bool 1
class | paper_arxiv_id stringlengths 10 13 | section_in_paper_id float64 1 294 ⌀ |
|---|---|---|---|---|---|
95,452 | \label{sec:IV}
In the following, we exhibit a list of gradient descent-based methods, which can be generalized to the Riemannian space and utilized in \gls{mo}mo strategy.
\subsection{GRADIENT DESCENT-BASED ALGORITHMS} \label{subsec:gradient}
Manifold-based optimization techniques, such as \gls{rn}rn, \gls{r... | GRADIENT DESCENT-BASED ALGORITHMS | false | 2507.06883 | 5 |
95,453 | \label{sec:V}
In the sequel, we present detailed applications of the \gls{mo}mo framework to power system optimization problems in \gls{epds}epds and \gls{epts}epts. Specifically, we demonstrate how different \gls{mo}mo techniques can be effectively employed for solving power flow. Through the application of suggeste... | MANIFOLDS IN POWER SYSTEMS OPTIMIZATION | false | 2507.06883 | 6 |
95,459 |
This research presents a full-stack toolchain comprising an automated scenario generation tool, a co-simulation platform, and a result analysis module. In accordance with the ISO 34502, three representative scenarios are selected from the highD dataset. The simulation results indicate that Apollo is capable of driving... | conclusion \& outlook | false | 2507.06884 | 4 |
95,464 |
\subsection{PRIVACY ANALYSIS}
Initially, local cumulant matrix $\{Cum^k\}$\{Cum^k\} are learned in the respective FL clients, which require the raw data of each client, and then are shared to the server for obtaining the global cumulant matrix $Cum$Cum. Next, The server utilizes the global cumulant matrix as the input... | Privacy and costs analysis | false | 2507.06888 | 5 |
860 |
Deep learning algorithms rely on large data sets for training a model to perform a complex task. However, annotating such large data sets usually requires a person to analyze each data point and label it accordingly, resulting in a time-consuming process. In particular, detecting particles in \ac{EM}EM images is extre... | Introduction | false | 2508.00563 | 1 |
861 |
\paragraph{Weakly supervised object detection.}Weakly supervised object detection. The requirement for fast annotation times is a long-standing problem in several fields, which has made \ac{WSOL}WSOL and \ac{WSOD}WSOD an active area of research in the last few years.\cite{oquab2015object} introduced a CNN architecture... | Related Work | false | 2508.00563 | 2 |
862 | \label{sec:methods}
Our method expects as input an \ac{EM}EM image, $I \in \R^{W \times H}$I \in \R^{W \times H}W \times H, the expected virus radius $r$r, and a classifier $C: \R^{W \times H} \rightarrow \R$C: \R^{W \times H}W \times H \rightarrow \R. We pre-train the classifier on image-level annotations, such that i... | Method | false | 2508.00563 | 3 |
863 |
In this section, we describe the experimental setup used to validate our method. To analyze our results on a variety of viruses, we have focused our experiments on the following five virus types: Herpes virus, Adeno virus, Noro virus, Papilloma virus, and Rota virus. Below, we will briefly discuss these with respect t... | Experiments | false | 2508.00563 | 4 |
1,320 |
We have investigated the low-rank \ac{ADI}ADI applied to the algebraic Lyapunov equation.
We empirically observed that simply performing certain parts of the \ac{ADI}ADI in a lower precision does not lead to highly accurate results.
We now revisit our hypotheses listed on \autopageref{list:hypotheses}list:hypotheses:... | Conclusions and future work | false | 2508.00722 | 4 |
3,030 | \label{sec:sat-in-monodic-c2}
In this section, we study the satisfiability problem of frugal \SmonCtwo and prove \NExpTime-completeness (which carries over to full \SmonCtwo), constituting the central result of the paper.
To get started, we provide an overview of the argument used to establish the result.
In \Cref{... | Satisfiability in Monodic Standpoint | false | 2508.00653 | 3 |
76,454 | \label{rel_work}
\subsection{Natural Language to Database Query Language}
Recent advances in large language models (LLMs) have significantly improved the ability to translate natural language into database query languages.
For instance, there has been extensive research on the Text2SQL task, which translates natur... | Related Work | false | 2505.05118 | 2 |
76,455 | \label{schema_filtering}
We now present schema filtering for Text2Cypher using a template from \cite{ozsoy2025text2cypher} (Table \ref{tab:instructions}), which includes instructions, schema, input question, and generated query fields. Our focus is on the schema field, where we experiment with two static and three dyna... | Schema Filtering for Text2Cypher | false | 2505.05118 | 3 |
76,456 | \label{experiment}
\subsection{Experimental Setup and Evaluation Metrics}
We conducted experiments using a publicly available Text2Cypher dataset \cite{ozsoy2025text2cypher}, focusing on a subset with accessible databases for query execution, resulting in 22,093 training and 2,471 test samples. Schema filtering was a... | Experimental Setup and Results | false | 2505.05118 | 4 |
76,463 |
\label{sec:experiments}
We evaluate the effectiveness of our approach with comprehensive experiments in various settings, including in-distribution performance analysis, large-scale (out-of-distribution) generalization analysis, real-world application, model components ablation study, and further analyses. We compa... | Experiments | false | 2505.05119 | 5 |
76,465 |
Option pricing is a fundamental problem in finance. Since the seminal work of \citet{black1973pricing}, numerous mathematical models and computational approaches have been developed to determine option prices. One common approach formulates the price of an option as the solution to a partial differential equation (PDE... | Introduction | false | 2505.05121 | 1 |
76,466 |
\label{sec:neural}
This section explains the two neural network methods used in this paper. Subsection \ref{sec:time} elaborates on the TDGF and Subsection \ref{sec:deep} on the DGM.
\subsection{Time Deep Gradient Flow Method}
\label{sec:time}
The TDGF is a neural network method to efficiently solve high-dimensional ... | Neural network methods | false | 2505.05121 | 2 |
95,465 |
In this section, we demonstrate that the situation can be easily extended to the Linear Gaussian Model. Suppose there are three variables $x$x, $y$y, and $z$z:
\begin{equation} \label{8}
\rho_{xy|z}=\frac{\rho_{xy}-\rho_{xz}\rho_{zy}}{\sqrt{(1-\rho_{xz}^2)(1-\rho_{zy}^2)}}
\end{equation}\begin{equation} \label{8}
\rh... | Extensions to More General Cases | false | 2507.06888 | 6 |
95,466 |
Power electronic converters in microgrids must be continuously monitored to guarantee safe operation. Recent studies have proposed noise-resilient residual networks [1], chaotic-DC arc-grounding detection [2], capsule-based compound-fault decoupling [3], and lightweight online classifiers for multiphase drives [4], il... | Introduction | false | 2507.06890 | 1 |
95,467 |
The Fractional-Order Memory-Enhanced Attack-Diagnosis Scheme (FO-MADS) framework is proposed to address fault and cyber-attack diagnosis in smart microgrids. Its core innovation lies in utilizing \textbf{single-sensor data} from the Point of Common Coupling (PCC), measuring bus voltage $V$V, active power $P$P, and rea... | Proposed FO-MADS Framework | false | 2507.06890 | 2 |
95,468 |
\subsection{PMR-AT Training Procedure}
The Progressive Memory-Replay Adversarial Training (PMR-AT) strategy is designed to enhance model robustness against evolving cyber-attack patterns. The specific training procedure is shown in Fig. 5. Unlike conventional adversarial training, PMR-AT incorporates four key innovat... | Progressive Memory-Replay Adversarial Training and Attack-Aware Loss Function | false | 2507.06890 | 3 |
95,469 |
\subsection{Experimental Setup}
1. \textbf{Dataset Generation}: A Simulink-based four-inverter microgrid testbed (Fig. 7) generated 5,600 VPQ waveform samples (50 Hz, 2 kHz sampling). Each sample spans a 0.2-s window centered at a fault instant ($t = 6.0$t = 6.0 s), yielding 400 points per window. The dataset include... | Simulation and Experimental Results | false | 2507.06890 | 4 |
95,470 |
This study introduces the Fractional-Order Memory-Enhanced Attack-Diagnosis Scheme (FO-MADS), a cost-effective framework that achieves cyber-physical resilience for smart microgrids, requiring only a single VPQ sensor at the point of common coupling. By jointly exploiting Caputo and Grünwald-Letnikov derivatives, a du... | Conclusion and Future Work | false | 2507.06890 | 5 |
95,471 |
\begin{figure}
\centering
\includegraphics[width=\linewidth]{figures/Main-3.pdf}
\vspace{-0.7cm}
\caption{\textbf{Efficiency-Performance Comparison for 1.5B Models (\textit{left}) and 7B Models (\textit{right}) in terms of Rollout Data Volume (i.e., total number of responses generated during training)... | Introduction | false | 2507.06892 | 1 |
95,477 |
To enable stable off-policy training, we adopt a V-trace~\citep{EspeholtSMSMWDF18Impala} formulation for generalized advantage estimation (GAE)~\citep{SchulmanMLJA15GAE}, which incorporates truncated importance sampling ratios to correct for policy mismatch. We first compute the temporal-difference error(TD-error) at... | Advantage Estimation | false | 2507.06892 | 7 |
95,478 |
\label{app:baseline_intro}
\subsection{1.5B Models}
\begin{itemize}
\item \textbf{Open-RS Series}~\citep{dang2025reinforcementopenrs}:
The Open-RS series employs the GRPO algorithm to train language models, using datasets constructed by filtering and combining existing corpora. Specifically, \textbf{Open-RS1} ... | A Brief Overview of Baseline Models | false | 2507.06892 | 8 |
95,479 |
\label{app:trainingde}
\paragraph{Hyperparmeters}Hyperparmeters The major hyperparameter choices are shown in Table~\ref{tab:hypers}.
\begin{table}[h]
\centering
\begin{tabular}{l|l}
\toprule
\textbf{Parameter} & \textbf{Value} \\
\midrule
\multicolumn{2}{l}{\textit{Training Configuration}} \\
\midrule
critic\_warm... | Training Details | false | 2507.06892 | 9 |
95,480 |
\label{app:training_curves}
\begin{figure}
\vspace{-0.2cm}
\centering
\includegraphics[width=\linewidth]{figures/min.pdf}
\vspace{-0.6cm}
\caption{\textbf{Training Efficiency Comparison for ReMix-PPO and PPO (1.5B) on Minerva.}\methodName achieves a score above 26\%, around \textbf{3x to 6x faster} th... | More Training Curves | false | 2507.06892 | 10 |
95,481 |
\label{app:case}
To better understand model’s reasoning behavior, we present a case study centered on a representative example that the base model is able to solve correctly. Figure compares the responses produced by three variants trained under distinct strategies: PPO, Mix-PPG, and Mix-PPG with an Increased UTD rat... | Case Study | false | 2507.06892 | 11 |
95,472 |
\label{subsec:preli}
\subsection{Reinforcement Learning for LLM Fine-tuning}
Reinforcement Fine-Tuning (RFT), also referred to as ReFT or RLVR (Reinforcement Learning with Verifiable Rewards), is a paradigm for adapting pre-trained LLMs to specific downstream tasks using RL~\citep{trung2024reft,jaech2024openai}. Un... | Preliminaries | false | 2507.06892 | 2 |
95,476 |
\label{sec:conclusion}
In this paper, we aim to address the notorious drawback of on-policy RFT methods (e.g., PPO and GRPO) on training inefficiency and prohibitive computational cost.
We launch the renaissance of
off-policy RL and propose Reincarnating Mix-policy Proximal Policy Gradient (\methodName), a general ap... | Conclusion | false | 2507.06892 | 6 |
95,482 |
\label{introduction}
AI evaluations have emerged as a valuable way to assess the capabilities and safety of large language models (LLMs). The UK AI Security Institute (UK AISI) identifies AI evaluations as a critical tool to advance our understanding of model capabilities and prevent catastrophic AI risks \yrcite{UKA... | Introduction | false | 2507.06893 | 1 |
76,468 |
\label{sec:implementation}
For the architecture we use the architecture from \citet{papapantoleon2024time} including the use of information about the option price in order to facilitate the training of the neural network:
\begin{align*}
X^1 & = \sigma_1 \left( W^1 \mathbf{x} + b^1 \right), \\
Z^l & = \sigma_1... | Implementation details | false | 2505.05121 | 4 |
76,470 |
\label{sec:conclusion}
This research analyzed the error of two neural network methods to solve option pricing PDEs: TDGF and DGM. We determined the empirical convergence rates of the $L^2$L^2-error of five parameters in both the Black--Scholes and the Heston model. We also considered the effect of these parameters on ... | Conclusion | false | 2505.05121 | 6 |
76,490 |
We denote by $\mathcal{N}_{r}(t)$\mathcal{N}_{r}r(t), a neighborhood in $C(t)$C(t) of radius $r\geq0$r\geq0 defined by
\begin{multline*}
\mathcal{N}_{r}(t):= \\
\{t^{\prime}\in\lbrack0,T]:|C_{t,t^{\prime}}|\leq
r\}\cup\{t^{\prime}\in\lbrack0,T]:|C_{t^{\prime},t}|\leq r\}.
\end{multline*}\begin{multline*}
\mathcal{N}... | Appendix. A basic vectorization algorithm based on the curvature of the silhouette contour | false | 2505.05132 | 7 |
76,491 |
Subglottic Stenosis (SGS) is a condition characterized by the narrowing of the airway in the subglottis, just below the vocal cords and before the trachea, which results in airway obstruction airflow restriction.
The percentage of narrowed airway or Stenosis Index (SI) is obtained by comparing the obstructed airway ... | Introduction | false | 2505.05136 | 1 |
95,484 |
Maintaining an open-source repository of AI evaluations introduces new challenges beyond traditional software development. Key challenges include:
\begin{itemize}
\item The rapid pace of new evaluation publications combined with the absence of centralized evaluation sources complicates discovery and curation efforts... | Learnings and Challenges | false | 2507.06893 | 3 |
268,031 |
In this paper, we explore interactive text-to-image person re-identification, which aims to improve the alignment between dynamic queries and challenging candidate images by leveraging external guidance from MLLMs. To achieve this, we develop an Interactive Cross-modal Learning (ICL) framework to alleviate the inheren... | Conclusion | false | 2506.11036 | 5 |
95,485 |
The challenges identified in maintaining $inspect\_evals$inspect\_evals point toward several promising research directions and infrastructure improvements that could benefit the broader AI evaluation community.
\textbf{Automated validation of evaluation implementations.} Current CI pipelines in $inspect\_evals$inspe... | Future Directions | false | 2507.06893 | 4 |
95,486 |
This paper documents the practical challenges and methodological insights gained from maintaining $inspect\_evals$inspect\_evals over eight months of active development.
The rapid pace of new evaluations, incomplete paper details, and buggy reference implementations create systematic barriers that individual organiz... | Conclusion | false | 2507.06893 | 5 |
95,487 |
This paper presents work whose goal is to advance the field of AI evaluations and safety. Our work on maintaining and improving AI evaluations has several potential societal consequences worth highlighting.
\textbf{Positive impacts.} By improving the quality, accessibility, and reliability of AI evaluations, this wo... | Impact Statement | false | 2507.06893 | 6 |
95,488 |
We thank JJ Allaire (UK AISI) and Charles Teague (RAND) for their mentorship and trust throughout the course of this project. We are grateful to Evan Miller (Anthropic) for his guidance on cross-model statistical testing, and to Jean-Stanislas Denain (Epoch AI) for his feedback on the $inspect\_evals$inspect\_evals d... | Acknowledgements | false | 2507.06893 | 7 |
139,043 |
This work was supported by the National Natural Science Foundation of China (No. 62401601).
\bibliographystyle{IEEEtran}IEEEtran
\bibliography{reference}
| Acknowledgment | false | 2504.10974 | 6 |
139,044 |
\IEEEPARstart{A}A{ccurate}ccurate medical image segmentation plays a pivotal role in lesion analysis and clinical decision-making. Colorectal cancer remains a global health challenge with high mortality rates, underscoring the critical need for early detection of precancerous polyps through endoscopic imaging. As the... | Introduction | false | 2504.10978 | 1 |
139,045 |
This section introduces AgentPolyp, an innovative intelligent system for polyp segmentation, as shown in Fig.~\ref{fig:2}, where the agent iteratively analyzes image degradation patterns, adaptively enhances input images, and improves performance based on segmentation strategies.
AgentPolyp performs complex analysis... | Methodology | false | 2504.10978 | 2 |
139,046 |
\subsection{Experimental Settings}
In this section, the performance of the AgentPolyp model is evaluated through comparative experiments, and U-Net~\cite{ronneberger2015u}, U-Net++~\cite{zhou2018unet++}, ResUNet-mod~\cite{zhang2018road }, ResUNet++~\cite{jha2019resunet++} and CTNet~\cite{10471227} are selected as the... | Experiments | false | 2504.10978 | 3 |
183,494 | \label{sec:DP_cyber_risks}
Although synthetic OPF datasets contribute to overall grid transparency and enable independent power flow analysis, they can also be misused by cyber adversaries launching attacks on the grid. One class of attacks, which is of interest to this work, is \textit{load redistribution} attacks. ... | Cyber Resilience Risks in Releasing Differentially Private OPF Datasets | false | 2503.14877 | 3 |
183,497 | \label{sec:conclusions}
We developed algorithms for synthesizing credible grid parameters from real-world systems for OPF analysis. Similar to existing DP algorithms, they obfuscate loads by injecting Laplacian noise and using post-processing; however, they differ in a post-processing stage which optimizes for the trad... | Conclusion | false | 2503.14877 | 6 |
95,489 |
Table~\ref{eval-costs} shows the estimated costs for running various evaluations across different models. These estimates highlight the significant cost differences between models, with GPT-4.5 being substantially more expensive than open-source alternatives like Llama-3.2-90B or commercial alternatives like DeepSeek... | Evaluation Cost Estimates | false | 2507.06893 | 8 |
95,490 |
\label{sec:intro}
Knowledge Graphs (KGs) play a crucial role in organizing and representing structured information across various domains, enhancing applications from information retrieval to complex question-answering systems \cite{KnowledgeGraphs,KGOppChall}. Enriching KGs to ensure they remain up-to-date requires l... | Introduction | false | 2507.06895 | 1 |
95,491 |
\label{sec:problem}
To formally define the problem of sentence-based multi-label RE, let us start by introducing the notion of annotated corpus and then define the task in terms of multi-label classification.
\begin{definition}[Annotated corpus]
Let us consider a reference KG denoted as $\mathcal{G} = (\mathcal{E}, \m... | Problem Definition | false | 2507.06895 | 2 |
95,492 |
\label{sec:model}
Figure \ref{fig:SCoRE_pipeline} illustrates the overall workflow of SCoRE. In the training phase:
\begin{enumerate}
\item each sentence in $\mathcal{C}_{train}^{\mathcal{G},\mathcal{S}'}$ goes through a single forward pass of a PLM encoder to get an hidden vector representation of its head and tail ... | SCoRE | false | 2507.06895 | 3 |
95,493 |
\label{sec:metrics}
In this section, we outline the evaluation criteria for assessing both the performance and environmental impact of sentence-level RE solutions. We start with commonly used RE metrics, explicitly defining their extensions to multi-label classification. Moreover, we illustrate the process we use to r... | Evaluation Metrics | false | 2507.06895 | 4 |
95,494 |
\label{sec:datasets}
Five annotated corpus were considered to conduct the experiments: four are well-established benchmarks for RE, while the fifth, Wiki20d, is our released FDS annotated corpus. Their features are summarized in Table \ref{sec:datasets_table}, detailing for each annotated corpus the reference KG, the ... | Datasets | false | 2507.06895 | 5 |
95,495 |
\label{sec:experiments}
\subsection{Settings Used in SOTA Models}
We compare SCoRE with four sentence-level RE methods and one bag-level RE method:
\begin{itemize}
%\item \textbf{RECON}\cite{recon}: a sentence-based RE approach that enhances context augmentation by leveraging a neural network to capture the conte... | Experimental Setup | false | 2507.06895 | 6 |
95,496 |
\label{sec:results}
\begin{table*}[t]
\centering
\caption{Performance comparison of SCoRE with state-of-the-art models using micro-F1 and macro-F1 metrics. Results from non-reproducible methods marked by an asterisk (*)}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\textbf{Model} &... | Results | false | 2507.06895 | 7 |
76,512 |
Spatially and sectorally detailed energy consumption and emissions data is a valuable resource for the energy
and climate research community. For instance, the energy consumption in road transport can support the planning of electric vehicle
charging infrastructure, facilitating the transition to electric mobility ... | Background \& Summary | false | 2505.05139 | 1 |
76,513 |
\subsection*{Overview}
The spatial disaggregation of FEC and emissions data is carried out in four steps, as illustrated in Figure \ref{fig:workflow_no_data_sources}. The following paragraphs provide an overview of these four steps, before going into more detail
in the sub-sections of this Methods section.
\textbf{... | Methods | false | 2505.05139 | 2 |
76,515 |
This study introduces a spatial disaggregation workflow that requires technical validation at two critical stages:
(1) the imputation of missing values in proxy data, and (2) the final disaggregation of FEC and emissions data.
The validation of the missing value imputation was shown already in the Methods section. ... | Technical Validation | false | 2505.05139 | 4 |
95,497 |
\label{sec:related}
Recent RE research has increasingly focused on deep learning solutions. Among those, non-PLM approaches often leverage alternative architectures and external knowledge resources, such as KgPool \cite{kgpool}, which enriches input context with KG facts for single-label extraction, and graph NN-based... | Related Works | false | 2507.06895 | 8 |
95,498 |
\label{sec:conclusion}
In this paper, we introduced SCoRE, a sentence-based multi-label RE tool designed to effectively handle the noise inherent in DS annotations. To the best of our knowledge SCoRE is the first approach that completely avoids finetuning, using the PLM solely as an informed prior during dataset creat... | Conclusions and Future Works | false | 2507.06895 | 9 |
95,499 |
Graphical User Interface (GUI) agents, powered by Large Vision-Language Models (LVLMs), have demonstrated remarkable capabilities in autonomously operating personal devices (e.g., desktop and mobile phone) or applications within them through visual understanding and interaction~\citep{cheng-etal-2024-seeclick,zhengG... | Introduction | false | 2507.06899 | 1 |
95,503 |
In this paper, we conducted the first study on backdoor attacks targeting the visual grounding capabilities of GUI agents. We introduce VisualTrap, a simple yet effective framework that poisons the fundamental grounding pre-training process of LVLMs with visual triggers, causing GUI agents to misinterpret interface e... | Conclusion | false | 2507.06899 | 5 |
95,504 |
Our research explores the vulnerability of GUI agents to backdoor attacks through visual grounding manipulation. While this work identifies significant security risks, our primary goal is to promote awareness about these vulnerabilities to strengthen the security of future agent systems before their widespread deploym... | Ethics Statement | false | 2507.06899 | 6 |
95,505 |
We thank the anonymous reviewers for their insightful comments. This research was also supported by the advanced computing resources provided by the Supercomputing Center of the USTC.
\bibliography{colm2025_conference}
\bibliographystyle{colm2025_conference}colm2025_conference
\newpage
| Acknowledgments | false | 2507.06899 | 7 |
95,500 |
$\bullet$\bullet \textbf{GUI Agent.}
Inspired by the great success of LLMs/LVLMs in various domains~\citep{gpt4,qwen,wang2024qwen2vlenhancingvisionlanguagemodels,pacar,nextquill,latentR}, GUI agents have evolved significantly in recent years, transitioning from early rule-based automation systems~\citep{ruleagent} ... | Related Work | false | 2507.06899 | 2 |
95,506 |
We identify several limitations in our work: 1) In the downstream fine-tuning of the end-to-end architecture, we assume that the user lacks sufficient resources to fully fine-tune the LVLM. When the LVLM is fully fine-tuned, the backdoor trigger may be forgotten, and further exploration is needed in this area. 2) Ou... | Limitation | false | 2507.06899 | 8 |
95,507 |
This section presents the details of the utilized datasets.
\subsection{Pretraining data}\label{sec:pretraining-apx}
Normal pertaining data. Following the SeeClick paper~\citep{cheng-etal-2024-seeclick}, we use diverse data to ensure robust grounding capabilities across different GUI contexts. Specifically, the pret... | Dataset Details | false | 2507.06899 | 9 |
95,501 |
\subsection{Formulation of GUI Agents}
\label{sec:agent-formulation}
GUI agents operate as autonomous systems designed to interact with graphical user interfaces through visual perception and action generation. These agents typically follow one of two architectural paradigms: a unified end-to-end approach or a mo... | Hijack Visual Grouding Ability of GUI Agents | false | 2507.06899 | 3 |
95,508 | \label{sec:more-case}
Figure~\ref{fig:more-case} shows additional 4 examples of case studies. As the figure shows, in each example, the agent conducted its action in the trigger position (red) instead of its initially intended position (green).
| Case Study | false | 2507.06899 | 10 |
95,509 | \label{sec:Visualization}
Figure~\ref{fig:trigger-size} shows the Visualization of Triggers with Varying Sizes ($5\times5$5\times5, $10\times10$10\times10, $20\times20$20\times20, and $50\times50$50\times50).
Figure~\ref{fig:trigger-intensity} shows the Visualization of Triggers with Varying Intensities (noisy intensit... | Visualization of Triggers with Varying Sizes and Intensities | false | 2507.06899 | 11 |
95,510 | \label{sec:appendix-defense}
Figure~\ref{fig:defense} presents the ASR after fine-tuning the LVLM with varying amounts of clean grounding data. The results are based on the pretraining phase evaluation setting.
The results show that when only the LLM component of LVLM is attacked, increasing fine-tuning data to 30\% o... | Defense Results | false | 2507.06899 | 12 |
95,511 |
The rapid proliferation of data streams from diverse sources, such as Internet of Things (IoT) devices, financial markets, social media platforms, and real-time video analytics, has underscored the critical need for efficient and adaptive data processing techniques~\cite{benczur2018online,assem2019machine}. Data strea... | Introduction | false | 2507.06901 | 1 |
95,512 |
\label{sec:related}
The field of data stream processing has seen significant advancements in adaptive sliding window techniques and the application of machine learning, including reinforcement learning, to address the challenges of dynamic and evolving data streams. This section provides a comprehensive review of the ... | Related Work | false | 2507.06901 | 2 |
95,513 |
\label{sec:method}
This section presents a comprehensive reinforcement learning (RL)-based framework for dynamically optimizing the sliding window size in multi-dimensional data streams. The approach formulates window size selection as an RL problem, where an agent learns to select the optimal window size based on the... | Methodology | false | 2507.06901 | 3 |
95,514 |
\label{sec:experiments}
This section evaluates the performance of our proposed RL-based adaptive sliding window method against state-of-the-art baselines on multiple benchmark datasets for multi-dimensional data streams. We compare our method in terms of classification accuracy, average window size, computational cost... | Experiments | false | 2507.06901 | 4 |
95,515 |
\label{sec:conclusion}
This paper presents a novel reinforcement learning-based approach for dynamically optimizing sliding window sizes in multi-dimensional data streams. By formulating the problem as an RL task, we enable an agent to learn an adaptive policy that balances historical context and responsiveness to rec... | Conclusion | false | 2507.06901 | 5 |
95,516 |
\label{sec:intro}
Humanoid robots, with their human-like morphology, represent a promising
paradigm for versatile robotic systems capable of operating in human-designed
environments. Recent years have witnessed remarkable advances in locomotion
~\cite{zhuangEmbraceCollisionsHumanoid2025,
zhangUnleashingHumanoidR... | INTRODUCTION | false | 2507.06905 | 1 |
95,517 |
\label{sec:related_work}
\subsection{Learning Legged Locomotion}
Reinforcement learning has emerged as the dominant paradigm for humanoid
locomotion control, demonstrating remarkable capabilities in learning complex
walking gaits and dynamic
behaviors~\cite{radosavovicLearningHumanoidLocomotion2024,
radosavovicH... | RELATED WORK | false | 2507.06905 | 2 |
95,518 |
\label{sec:problem_formulation}
We formulate the humanoid loco-manipulation task as a goal-conditioned Markov
Decision Process (MDP) defined by the tuple $\mathcal{M} = \langle \mathcal{S},
\mathcal{A}, \mathcal{G}, P, R, \gamma \rangle$\mathcal{M} = \langle \mathcal{S},
\mathcal{A}, \mathcal{G}, P, R, \gamma... | PROBLEM FORMULATION | false | 2507.06905 | 3 |
144 |
The need to broaden early engagement with quantum information science requires more than simply exposing students to quantum concepts; it calls for thoughtfully designed, theoretically grounded learning environments that support deep understanding of ideas that challenge even expert intuition. A promising path forward... | Conclusion | false | 2508.00668 | 8 |
1,322 |
\label{sec:intro}
Active matter theories describe the physics of non-equilibrium systems in which the scale of the drive is smaller than the minimal lengthscale at which the theoretical description is valid and the driving mechanism is not spatially correlated. That is, unlike traditional non-equilibrium systems in w... | Introduction | false | 2508.00684 | 1 |
1,323 |
\label{sec:screw_symmetry}
Chiral mechanical forces have been identified in a range of continuum materials and mesophases~\cite{fruchart2023,maitra2025}. A basic example is the odd elasticity of two-dimensional isotropic solids, which support a chiral force density $K_o \,{\bf J} \cdot \nabla^2 {\bf u}$K_o \,{\bf J}\... | Chiral Mechanics and Screw Symmetry in Cholesterics | false | 2508.00684 | 2 |
76,520 |
The authors declare no competing interests.
| Competing interests | false | 2505.05139 | 9 |
76,521 |
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{figures/workflow_no_data_sources.pdf}
\caption{The steps involved in the spatial disaggregation of emissions and Final Energy Consumption (FEC) data
from country (NUTS0) to municipal (LAU) level.}
\label{fig:workflow_no_data_sources}... | Figures \& Tables | false | 2505.05139 | 10 |
95,519 |
\label{sec:ulc}
\begin{figure*}[t]
\centering
\includegraphics[width=1.0\textwidth]{imgs/method.png}
\caption{Method overview of the \textcolor{myred}{U}nified
\textcolor{myred}{L}oco-Manipulation \textcolor{myred}{C}ontroller
(\ULC). Our approach employs massively parallel reinforcement le... | U}NIFIED {L}OCO-MANIPULATION {C}ONTROL | false | 2507.06905 | 4 |
95,520 |
\label{sec:teleop}
\begin{figure*}[t]
\centering
\includegraphics[width=1.0\textwidth]{imgs/tele.png}
\caption{Schematic of the teleoperation system. The system acquires multimodal inputs from VR headsets and remote controllers, and processes them through custom algorithms to generate structured robot cont... | TELEOPERATION SYSTEM FOR | false | 2507.06905 | 5 |
95,521 |
\label{sec:exp}
\begin{figure*}[htbp]
\centering
\includegraphics[width=1.0\textwidth]{imgs/tracking_ranges_heatmap.pdf}
\caption{Comparison of reachable command ranges for different methods. Values show operational limits for root height (m), orientations (rad), and arm control capabilities.}
\label{f... | EXPERIMENT | false | 2507.06905 | 6 |
95,522 |
\label{sec:conclusion}
We presented \textcolor{myred}{ULC}, a unified controller for humanoid
loco-manipulation that, to the best of our knowledge, is the first to
simultaneously achieve unified whole-body control, large operational workspace,
and high-precision tracking. By integrating all degrees of freedom in a sin... | CONCLUSIONS and LIMITATIONS | false | 2507.06905 | 7 |
95,523 |
\label{sec:appendix}
\subsection{Communication Architecture}
\label{subsec:append_communication}
The teleoperation system employs a distributed communication architecture that
coordinates multiple components across different computational nodes. Dual
cameras mounted on the robot's onboard computer transmit stereo ima... | APPENDIX | false | 2507.06905 | 8 |
95,530 | \label{sec:conclusion}
In this study, we investigated various voting strategies within an NVML system designed for safety-critical traffic sign recognition. By developing a safety metric grounded in the FMEA method, we were able to assess system safety and assign safety-aware weights for both weighted voting and w... | Conclusion | false | 2507.06907 | 7 |
95,524 |
\label{sec:intro}
In recent years, machine learning (ML) has become a fundamental technology in safety-critical domains, including autonomous driving. In particular, computer vision (CV) systems powered by ML models play a crucial role in enabling autonomous vehicles to perceive and interpret their surroundings \cite... | Introduction | false | 2507.06907 | 1 |
95,525 | \label{sec:related_work}
Recent advancements in multi-version ML techniques have aimed at enhancing the reliability of ML systems.
Xu et al. explored the potential of developing fault-tolerant deep learning systems through model redundancy (i.e., NV-DNN \citep{xu2019nv}). Experimental results indicate that their appr... | Related work | false | 2507.06907 | 2 |
95,526 | \label{sec:background}
\subsection{Traffic sign recognition}
For autonomous vehicles, the precise identification of traffic signs is a cornerstone of safe operation. These vehicles rely on systems that can rapidly detect and interpret regulatory, warning, and informational signs to follow road rules and maintain ... | Background | false | 2507.06907 | 3 |
95,527 | \label{sec:voting_mechanism}
The NVML framework aggregates predictions from several machine learning models. A robust voting scheme is essential to fuse these individual predictions into a trustworthy final decision. In our exploration of traffic sign recognition, we evaluated various voting strategies. Depending on... | Voting mechanism | false | 2507.06907 | 4 |
95,533 |
\jp{
The emergence of \acp{LLM} has significantly accelerated research across a range of critical and specialized domains, including
psychotherapy~\cite{sun2024eliciting,sun2025rethinking,sun2024script},
medicine~\cite{yan2022remedi,wang2023pre}, and
industrial coaching~\cite{pei2024autonomous}. In the legal field, ... | Introduction | false | 2507.06909 | 1 |
95,534 |
\Acf{LJP}LJP is a critical task for smart legal assistants, which aims to predict the outcomes of legal cases given the description of facts~\cite{cui2022survey}.
These outcomes usually include the types of the charge(s) and terms of penalty.
Due to differences in national legal systems~\cite{sznycer2020origins}.
Spe... | Related Work | false | 2507.06909 | 2 |
95,535 |
\subsection{Raw data collection} We constructed the MPMCP dataset using first-instance documents collected from China Judgments Online\footnote{\url{https://wenshu.court.gov.cn/}}, covering the period from 1998 to 2021.
\jp{We exclusively obtain criminal cases with judgment outcomes and retain documents that clearly i... | Dataset Construction | false | 2507.06909 | 3 |
183,504 |
This research is supported by
the National Science Fund for Distinguished Young Scholars (62125601),
the Beijing Natural Science Foundation (IS23060),
the National Natural Science Foundation of China (62172035),
the Youth Teacher International Exchange \& Growth Program (No. QNXM20250001),
and the Fundamental Resear... | Acknowledgments | false | 2503.14880 | 7 |
95,540 |
\label{sec:prompt-template}
Prompt templates for \acp{LLM}LLM to generate outcomes with an example or without an example are shown in Table~\ref{tab:prompt template}.
\begin{CJK}{UTF8}{gbsn} % UTF8 encoding and gbsn (Simplified Chinese) font
\input{Content/Tables/prompt_templates}
\end{CJK}\begin{CJK}{UTF8}{gbsn} % U... | Prompt template | false | 2507.06909 | 8 |
95,536 |
\subsection{Benchmark Models}
We leverage the following five prevailing open-source \acp{LLM}LLM for Chinese \ac{LJP}LJP as benchmark models to generate outputs in four scenarios.
\noindent \textbf{MT5}~\cite{xue2021mt5}, a T5 variant with multilingual capabilities, pre-trained on a novel dataset derived from Co... | Experimental Setup | false | 2507.06909 | 4 |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.