Background: Recruitment for cohorts involving complex liver diseases, such as hepatocellular carcinoma and liver cirrhosis, often requires interpreting semantically complex criteria. Traditional manual screening methods are time-consuming and prone to errors. While AI-powered pre-screening offers potential solutions, challenges remain regarding accuracy, efficiency, and data privacy. Methods: We developed a novel patient pre-screening pipeline that leverages clinical expertise to guide the precise, safe, and efficient application of large language models. The pipeline breaks down complex criteria into a series of composite questions and then employs two strategies to perform semantic question-answering through electronic health records - (1) Pathway A, Anthropomorphized Experts' Chain of Thought strategy, and (2) Pathway B, Preset Stances within an Agent Collaboration strategy, particularly in managing complex clinical reasoning scenarios. The pipeline is evaluated on three key metrics-precision, time consumption, and counterfactual inference - at both the question and criterion levels. Results: Our pipeline achieved high precision (0.921, in criteria level) and efficiency (0.44s per task). Pathway B excelled in complex reasoning, while Pathway A was effective in precise data extraction with faster processing times. Both pathways achieved comparable precision. The pipeline showed promising results in hepatocellular carcinoma (0.878) and cirrhosis trials (0.843). Conclusions: This data-secure and time-efficient pipeline shows high precision in hepatopathy trials, providing promising solutions for streamlining clinical trial workflows. Its efficiency and adaptability make it suitable for improving patient recruitment. And its capability to function in resource-constrained environments further enhances its utility in clinical settings.

Computer vision is a critical component in a wide range of real-world applications, including plant monitoring in agriculture and handwriting classification in digital systems. However, developing high-performance computer vision models traditionally demands both machine learning (ML) expertise and domain-specific knowledge, making the process costly, labor-intensive, and inaccessible to many. Large language model (LLM) agents have emerged as a promising solution to automate this workflow, but most existing methods share a common limitation: they attempt to optimize entire pipelines in a single step before evaluation, making it difficult to attribute improvements to specific changes. This lack of granularity leads to unstable optimization and slower convergence, limiting their effectiveness. To address this, we introduce Iterative Refinement, a novel strategy for LLM-driven ML pipeline design inspired by how human ML experts iteratively refine models, focusing on one component at a time rather than making sweeping changes all at once. By systematically updating individual components based on real training feedback, Iterative Refinement improves stability, interpretability, and overall model performance. We implement this strategy in IMPROVE, an end-to-end LLM agent framework for automating and optimizing object classification pipelines. Through extensive evaluations across datasets of varying sizes and domains, including standard benchmarks and Kaggle competition datasets, we demonstrate that Iterative Refinement enables IMPROVE to consistently achieve better performance over existing zero-shot LLM-based approaches. These findings establish Iterative Refinement as an effective new strategy for LLM-driven ML automation and position IMPROVE as an accessible solution for building high-quality computer vision models without requiring ML expertise.

The complete automation of cyber-attacks is an area of growing interest since the surge of Large Language Models (LLMs) in recent years. Although the application of LLM in all areas of cybersecurity has flourished, the creation of attacking LLM agents that can act independently is among the most popular options [1]. Attacking LLM agents can perform automatic security testing of applications, lowering the cost for organizations to find vulnerabilities and misconfiguration problems and identify other security issues [2]. Existing automated attacking agents, such as PenHeal [2], AutoAttacker [3], and HackSynth [4] show promising results but with clear limitations. Agents are unable to work so far without occasional mistakes and hallucinations.

Recent advancements in software engineering (SWE) agents have largely followed a $\textit{tool-based paradigm}$, where agents interact with hand-engineered tool APIs to perform specific tasks. While effective for specialized tasks, these methods fundamentally lack generalization, as they require predefined tools for each task and do not scale across programming languages and domains. We introduce $\texttt{Programming with Pixels}$ (PwP), an agent environment that unifies software development tasks by enabling $\textit{computer-use agents}$-agents that operate directly within an IDE through visual perception, typing, and clicking, rather than relying on predefined tool APIs. To systematically evaluate these agents, we propose $\texttt{PwP-Bench}$, a benchmark that unifies existing SWE benchmarks spanning tasks across multiple programming languages, modalities, and domains under a task-agnostic state and action space. Our experiments demonstrate that general-purpose computer-use agents can approach or even surpass specialized tool-based agents on a variety of SWE tasks without the need for hand-engineered tools. However, our analysis shows that current models suffer from limited visual grounding and fail to exploit many IDE tools that could simplify their tasks. When agents can directly access IDE tools, without visual interaction, they show significant performance improvements, highlighting the untapped potential of leveraging built-in IDE capabilities. Our results establish PwP as a scalable testbed for building and evaluating the next wave of software engineering agents. We release code and data at https://programmingwithpixels.com

English as a Second Language (ESL) speakers often struggle to engage in group discussions due to language barriers. While moderators can facilitate participation, few studies assess conversational engagement and evaluate moderation effectiveness. To address this gap, we develop a dataset comprising 17 sessions from an online ESL conversation club, which includes both moderated and non-moderated discussions. We then introduce an approach that integrates automatic ESL dialogue assessment and a framework that categorizes moderation strategies. Our findings indicate that moderators help improve the flow of topics and start/end a conversation. Interestingly, we find active acknowledgement and encouragement to be the most effective moderation strategy, while excessive information and opinion sharing by moderators has a negative impact. Ultimately, our study paves the way for analyzing ESL group discussions and the role of moderators in non-native conversation settings.

Large language models (LLMs) have significantly advanced formal theorem proving, yet the scarcity of high-quality training data constrains their capabilities in complex mathematical domains. Combinatorics, a cornerstone of mathematics, provides essential tools for analyzing discrete structures and solving optimization problems. However, its inherent complexity makes it particularly challenging for automated theorem proving (ATP) for combinatorial identities. To address this, we manually construct LeanComb, combinatorial identities benchmark in Lean, which is, to our knowledge, the first formalized theorem proving benchmark built for combinatorial identities. We develop an Automated Theorem Generator for Combinatorial Identities, ATG4CI, which combines candidate tactics suggested by a self-improving large language model with a Reinforcement Learning Tree Search approach for tactic prediction. By utilizing ATG4CI, we generate a LeanComb-Enhanced dataset comprising 260K combinatorial identities theorems, each with a complete formal proof in Lean, and experimental evaluations demonstrate that models trained on this dataset can generate more effective tactics, thereby improving success rates in automated theorem proving for combinatorial identities.

This paper argues that generating output tokens is more effective than using pooled representations for prediction tasks because token-level generation retains more mutual information. Since LLMs are trained on massive text corpora using next-token prediction, generation aligns naturally with their learned behavior. Using the Data Processing Inequality (DPI), we provide both theoretical and empirical evidence supporting this claim. However, autoregressive models face two key challenges when used for prediction: (1) exposure bias, where the model sees ground truth tokens during training but relies on its own predictions during inference, leading to errors, and (2) format mismatch, where discrete tokens do not always align with the tasks required output structure. To address these challenges, we introduce PredGen(Predicting Through Generating), an end to end framework that (i) uses scheduled sampling to reduce exposure bias, and (ii) introduces a task adapter to convert the generated tokens into structured outputs. Additionally, we introduce Writer-Director Alignment Loss (WDAL), which ensures consistency between token generation and final task predictions, improving both text coherence and numerical accuracy. We evaluate PredGen on multiple classification and regression benchmarks. Our results show that PredGen consistently outperforms standard baselines, demonstrating its effectiveness in structured prediction tasks.

Large language models (LLMs) have been increasingly used in time series analysis. However, the potential of multimodal LLMs (MLLMs), particularly vision-language models, for time series remains largely under-explored. One natural way for humans to detect time series anomalies is through visualization and textual description. Motivated by this, we raise a critical and practical research question: Can multimodal LLMs perform time series anomaly detection? To answer this, we propose VisualTimeAnomaly benchmark to evaluate MLLMs in time series anomaly detection (TSAD). Our approach transforms time series numerical data into the image format and feed these images into various MLLMs, including proprietary models (GPT-4o and Gemini-1.5) and open-source models (LLaVA-NeXT and Qwen2-VL), each with one larger and one smaller variant. In total, VisualTimeAnomaly contains 12.4k time series images spanning 3 scenarios and 3 anomaly granularities with 9 anomaly types across 8 MLLMs. Starting with the univariate case (point- and range-wise anomalies), we extend our evaluation to more practical scenarios, including multivariate and irregular time series scenarios, and variate-wise anomalies. Our study reveals several key insights: 1) MLLMs detect range- and variate-wise anomalies more effectively than point-wise anomalies. 2) MLLMs are highly robust to irregular time series, even with 25% of the data missing. 3) Open-source MLLMs perform comparably to proprietary models in TSAD. While open-source MLLMs excel on univariate time series, proprietary MLLMs demonstrate superior effectiveness on multivariate time series. To the best of our knowledge, this is the first work to comprehensively investigate MLLMs for TSAD, particularly for multivariate and irregular time series scenarios. We release our dataset and code at https://github.com/mllm-ts/VisualTimeAnomaly to support future research.

We present DocPuzzle, a rigorously constructed benchmark for evaluating long-context reasoning capabilities in large language models (LLMs). This benchmark comprises 100 expert-level QA problems requiring multi-step reasoning over long real-world documents. To ensure the task quality and complexity, we implement a human-AI collaborative annotation-validation pipeline. DocPuzzle introduces an innovative evaluation framework that mitigates guessing bias through checklist-guided process analysis, establishing new standards for assessing reasoning capacities in LLMs. Our evaluation results show that: 1)Advanced slow-thinking reasoning models like o1-preview(69.7%) and DeepSeek-R1(66.3%) significantly outperform best general instruct models like Claude 3.5 Sonnet(57.7%); 2)Distilled reasoning models like DeepSeek-R1-Distill-Qwen-32B(41.3%) falls far behind the teacher model, suggesting challenges to maintain the generalization of reasoning capabilities relying solely on distillation.

Large Language Models (LLMs) have demonstrated strong reasoning capabilities across various tasks. However, even minor variations in query phrasing, despite preserving the underlying semantic meaning, can significantly affect their performance. To address this, we focus on enhancing LLMs' awareness of symmetry in query variations and propose syMmetry-ENhanceD (MEND) Data Augmentation, a data-centric approach that improves the model's ability to extract useful information from context. Unlike existing methods that emphasize reasoning chain augmentation, our approach improves model robustness at the knowledge extraction stage through query augmentations, enabling more data-efficient training and stronger generalization to Out-of-Distribution (OOD) settings. Extensive experiments on both logical and arithmetic reasoning tasks show that MEND enhances reasoning performance across diverse query variations, providing new insight into improving LLM robustness through structured dataset curation.