De-identification in the healthcare setting is an application of NLP where automated algorithms are used to remove personally identifying information of patients (and, sometimes, providers). With the recent rise of generative large language models (LLMs), there has been a corresponding rise in the number of papers that apply LLMs to de-identification. Although these approaches often report near-perfect results, significant challenges concerning reproducibility and utility of the research papers persist. This paper identifies three key limitations in the current literature: inconsistent reporting metrics hindering direct comparisons, the inadequacy of traditional classification metrics in capturing errors which LLMs may be more prone to (i.e., altering clinically relevant information), and lack of manual validation of automated metrics which aim to quantify these errors. To address these issues, we first present a survey of LLM-based de-identification research, highlighting the heterogeneity in reporting standards. Second, we evaluated a diverse set of models to quantify the extent of inappropriate removal of clinical information. Next, we conduct a manual validation of an existing evaluation metric to measure the removal of clinical information, employing clinical experts to assess their efficacy. We highlight poor performance and describe the inherent limitations of such metrics in identifying clinically significant changes. Lastly, we propose a novel methodology for the detection of clinically relevant information removal.

Benchmarks such as SWE-bench and ARC-AGI demonstrate how shared datasets accelerate progress toward artificial general intelligence (AGI). We introduce VCBench, the first benchmark for predicting founder success in venture capital (VC), a domain where signals are sparse, outcomes are uncertain, and even top investors perform modestly. At inception, the market index achieves a precision of 1.9%. Y Combinator outperforms the index by a factor of 1.7x, while tier-1 firms are 2.9x better. VCBench provides 9,000 anonymized founder profiles, standardized to preserve predictive features while resisting identity leakage, with adversarial tests showing more than 90% reduction in re-identification risk. We evaluate nine state-of-the-art large language models (LLMs). DeepSeek-V3 delivers over six times the baseline precision, GPT-4o achieves the highest F0.5, and most models surpass human benchmarks. Designed as a public and evolving resource available at vcbench.com, VCBench establishes a community-driven standard for reproducible and privacy-preserving evaluation of AGI in early-stage venture forecasting.

Web agents powered by large language models (LLMs) can autonomously perform complex, multistep tasks in dynamic web environments. However, current evaluations mostly focus on the overall success while overlooking intermediate errors. This limits insight into failure modes and hinders systematic improvement. This work analyzes existing benchmarks and highlights the lack of fine-grained diagnostic tools. To address this gap, we propose a modular evaluation framework that decomposes agent pipelines into interpretable stages for detailed error analysis. Using the SeeAct framework and the Mind2Web dataset as a case study, we show how this approach reveals actionable weaknesses missed by standard metrics - paving the way for more robust and generalizable web agents.

Safety alignment aims to prevent Large Language Models (LLMs) from responding to harmful queries. To strengthen safety protections, jailbreak methods are developed to simulate malicious attacks and uncover vulnerabilities. In this paper, we introduce HILL (Hiding Intention by Learning from LLMs), a novel jailbreak approach that systematically transforms imperative harmful requests into learning-style questions with only straightforward hypotheticality indicators. Further, we introduce two new metrics to thoroughly evaluate the utility of jailbreak methods. Experiments on the AdvBench dataset across a wide range of models demonstrate HILL's strong effectiveness, generalizability, and harmfulness. It achieves top attack success rates on the majority of models and across malicious categories while maintaining high efficiency with concise prompts. Results of various defense methods show the robustness of HILL, with most defenses having mediocre effects or even increasing the attack success rates. Moreover, the assessment on our constructed safe prompts reveals inherent limitations of LLMs' safety mechanisms and flaws in defense methods.

This paper documents early research conducted in 2022 on defending against prompt injection attacks in large language models, providing historical context for the evolution of this critical security domain. This research focuses on two adversarial attacks against Large Language Models (LLMs): prompt injection and goal hijacking. We examine how to construct these attacks, test them on various LLMs, and compare their effectiveness. We propose and evaluate a novel defense technique called Adversarial Fine-Tuning. Our results show that, without this defense, the attacks succeeded 31\% of the time on GPT-3 series models. When using our Adversarial Fine-Tuning approach, attack success rates were reduced to near zero for smaller GPT-3 variants (Ada, Babbage, Curie), though we note that subsequent research has revealed limitations of fine-tuning-based defenses. We also find that more flexible models exhibit greater vulnerability to these attacks. Consequently, large models such as GPT-3 Davinci are more vulnerable than smaller models like GPT-2. While the specific models tested are now superseded, the core methodology and empirical findings contributed to the foundation of modern prompt injection defense research, including instruction hierarchy systems and constitutional AI approaches.

The rapid advancement of large language models (LLMs) has drawn urgent attention to the task of machine-generated text detection (MGTD). However, existing approaches struggle in complex real-world scenarios: zero-shot detectors rely heavily on scoring model's output distribution while training-based detectors are often constrained by overfitting to the training data, limiting generalization. We found that the performance bottleneck of training-based detectors stems from the misalignment between training objective and task needs. To address this, we propose Direct Discrepancy Learning (DDL), a novel optimization strategy that directly optimizes the detector with task-oriented knowledge. DDL enables the detector to better capture the core semantics of the detection task, thereby enhancing both robustness and generalization. Built upon this, we introduce DetectAnyLLM, a unified detection framework that achieves state-of-the-art MGTD performance across diverse LLMs. To ensure a reliable evaluation, we construct MIRAGE, the most diverse multi-task MGTD benchmark. MIRAGE samples human-written texts from 10 corpora across 5 text-domains, which are then re-generated or revised using 17 cutting-edge LLMs, covering a wide spectrum of proprietary models and textual styles. Extensive experiments on MIRAGE reveal the limitations of existing methods in complex environment. In contrast, DetectAnyLLM consistently outperforms them, achieving over a 70% performance improvement under the same training data and base scoring model, underscoring the effectiveness of our DDL. Project page: {https://fjc2005.github.io/detectanyllm}.

E-Commerce customer support requires quick and accurate answers grounded in product data and past support cases. This paper develops a novel retrieval-augmented generation (RAG) framework that uses knowledge graphs (KGs) to improve the relevance of the answer and the factual grounding. We examine recent advances in knowledge-augmented RAG and chatbots based on large language models (LLM) in customer support, including Microsoft's GraphRAG and hybrid retrieval architectures. We then propose a new answer synthesis algorithm that combines structured subgraphs from a domain-specific KG with text documents retrieved from support archives, producing more coherent and grounded responses. We detail the architecture and knowledge flow of our system, provide comprehensive experimental evaluation, and justify its design in real-time support settings. Our implementation demonstrates 23\% improvement in factual accuracy and 89\% user satisfaction in e-Commerce QA scenarios.

The proliferation of hate speech on social media necessitates automated detection systems that balance accuracy with computational efficiency. This study evaluates 38 model configurations in detecting hate speech across datasets ranging from 6.5K to 451K samples. We analyze transformer architectures (e.g., BERT, RoBERTa, Distil-BERT), deep neural networks (e.g., CNN, LSTM, GRU, Hierarchical Attention Networks), and traditional machine learning methods (e.g., SVM, CatBoost, Random Forest). Our results show that transformers, particularly RoBERTa, consistently achieve superior performance with accuracy and F1-scores exceeding 90%. Among deep learning approaches, Hierarchical Attention Networks yield the best results, while traditional methods like CatBoost and SVM remain competitive, achieving F1-scores above 88% with significantly lower computational costs. Additionally, our analysis highlights the importance of dataset characteristics, with balanced, moderately sized unprocessed datasets outperforming larger, preprocessed datasets. These findings offer valuable insights for developing efficient and effective hate speech detection systems.

Automated kernel design is critical for overcoming software ecosystem barriers in emerging hardware platforms like RISC-V. While large language models (LLMs) have shown promise for automated kernel optimization, demonstrating success in CUDA domains with comprehensive technical documents and mature codebases, their effectiveness remains unproven for reference-scarce domains like RISC-V. We present Evolution of Kernels (EoK), a novel LLM-based evolutionary program search framework that automates kernel design for domains with limited reference material. EoK mitigates reference scarcity by mining and formalizing reusable optimization ideas (general design principles + actionable thoughts) from established kernel libraries' development histories; it then guides parallel LLM explorations using these ideas, enriched via Retrieval-Augmented Generation (RAG) with RISC-V-specific context, prioritizing historically effective techniques. Empirically, EoK achieves a median 1.27x speedup, surpassing human experts on all 80 evaluated kernel design tasks and improving upon prior LLM-based automated kernel design methods by 20%. These results underscore the viability of incorporating human experience into emerging domains and highlight the immense potential of LLM-based automated kernel optimization.

Online toxic content has grown into a pervasive phenomenon, intensifying during times of crisis, elections, and social unrest. A significant amount of research has been focused on detecting or analyzing toxic content using machine-learning approaches. The proliferation of toxic content across digital platforms has spurred extensive research into automated detection mechanisms, primarily driven by advances in machine learning and natural language processing. Overall, the present study represents the synthesis of 140 publications on different types of toxic content on digital platforms. We present a comprehensive overview of the datasets used in previous studies focusing on definitions, data sources, challenges, and machine learning approaches employed in detecting online toxicity, such as hate speech, offensive language, and harmful discourse. The dataset encompasses content in 32 languages, covering topics such as elections, spontaneous events, and crises. We examine the possibility of using existing cross-platform data to improve the performance of classification models. We present the recommendations and guidelines for new research on online toxic consent and the use of content moderation for mitigation. Finally, we present some practical guidelines to mitigate toxic content from online platforms.