In this work-in-progress, we investigate the certification of AI systems, focusing on the practical application and limitations of existing certification catalogues in the light of the AI Act by attempting to certify a publicly available AI system. We aim to evaluate how well current approaches work to effectively certify an AI system, and how publicly accessible AI systems, that might not be actively maintained or initially intended for certification, can be selected and used for a sample certification process. Our methodology involves leveraging the Fraunhofer AI Assessment Catalogue as a comprehensive tool to systematically assess an AI model's compliance with certification standards. We find that while the catalogue effectively structures the evaluation process, it can also be cumbersome and time-consuming to use. We observe the limitations of an AI system that has no active development team anymore and highlighted the importance of complete system documentation. Finally, we identify some limitations of the certification catalogues used and proposed ideas on how to streamline the certification process.

Machine Learning models thrive on vast datasets, continuously adapting to provide accurate predictions and recommendations. However, in an era dominated by privacy concerns, Machine Unlearning emerges as a transformative approach, enabling the selective removal of data from trained models. This paper examines methods such as Naive Retraining and Exact Unlearning via the SISA framework, evaluating their Computational Costs, Consistency, and feasibility using the $\texttt{HSpam14}$ dataset. We explore the potential of integrating unlearning principles into Positive Unlabeled (PU) Learning to address challenges posed by partially labeled datasets. Our findings highlight the promise of unlearning frameworks like $\textit{DaRE}$ for ensuring privacy compliance while maintaining model performance, albeit with significant computational trade-offs. This study underscores the importance of Machine Unlearning in achieving ethical AI and fostering trust in data-driven systems.

This paper analyzes the safety of Large Language Models (LLMs) in interactions with children below age of 18 years. Despite the transformative applications of LLMs in various aspects of children's lives such as education and therapy, there remains a significant gap in understanding and mitigating potential content harms specific to this demographic. The study acknowledges the diverse nature of children often overlooked by standard safety evaluations and proposes a comprehensive approach to evaluating LLM safety specifically for children. We list down potential risks that children may encounter when using LLM powered applications. Additionally we develop Child User Models that reflect the varied personalities and interests of children informed by literature in child care and psychology. These user models aim to bridge the existing gap in child safety literature across various fields. We utilize Child User Models to evaluate the safety of six state of the art LLMs. Our observations reveal significant safety gaps in LLMs particularly in categories harmful to children but not adults

Multimodal Large Language Models (MLLMs) have serious security vulnerabilities.While safety alignment using multimodal datasets consisting of text and data of additional modalities can effectively enhance MLLM's security, it is costly to construct these datasets. Existing low-resource security alignment methods, including textual alignment, have been found to struggle with the security risks posed by additional modalities. To address this, we propose Synthetic Embedding augmented safety Alignment (SEA), which optimizes embeddings of additional modality through gradient updates to expand textual datasets. This enables multimodal safety alignment training even when only textual data is available. Extensive experiments on image, video, and audio-based MLLMs demonstrate that SEA can synthesize a high-quality embedding on a single RTX3090 GPU within 24 seconds. SEA significantly improves the security of MLLMs when faced with threats from additional modalities. To assess the security risks introduced by video and audio, we also introduced a new benchmark called VA-SafetyBench. High attack success rates across multiple MLLMs validate its challenge. Our code and data will be available at https://github.com/ZeroNLP/SEA.

The recent advances in large language models (LLMs) have revolutionized industries such as finance, marketing, and customer service by enabling sophisticated natural language processing tasks. However, the broad adoption of LLMs brings significant challenges, particularly in the form of social biases that can be embedded within their outputs. Biases related to gender, age, and other sensitive attributes can lead to unfair treatment, raising ethical concerns and risking both company reputation and customer trust. This study examined bias in finance-related marketing slogans generated by LLMs (i.e., ChatGPT) by prompting tailored ads targeting five demographic categories: gender, marital status, age, income level, and education level. A total of 1,700 slogans were generated for 17 unique demographic groups, and key terms were categorized into four thematic groups: empowerment, financial, benefits and features, and personalization. Bias was systematically assessed using relative bias calculations and statistically tested with the Kolmogorov-Smirnov (KS) test against general slogans generated for any individual. Results revealed that marketing slogans are not neutral; rather, they emphasize different themes based on demographic factors. Women, younger individuals, low-income earners, and those with lower education levels receive more distinct messaging compared to older, higher-income, and highly educated individuals. This underscores the need to consider demographic-based biases in AI-generated marketing strategies and their broader societal implications. The findings of this study provide a roadmap for developing more equitable AI systems, highlighting the need for ongoing bias detection and mitigation efforts in LLMs.

Preference alignment via reward models helps build safe, helpful, and reliable large language models (LLMs). However, subjectivity in preference judgments and the lack of representative sampling in preference data collection can introduce new biases, hindering reward models' fairness and equity. In this work, we introduce a framework for evaluating dialect biases in reward models and conduct a case study on biases against African American Language (AAL) through several experiments comparing reward model preferences and behavior on paired White Mainstream English (WME) and both machine-translated and human-written AAL corpora. We show that reward models are less aligned with human preferences when processing AAL texts vs. WME ones (-4\% accuracy on average), frequently disprefer AAL-aligned texts vs. WME-aligned ones, and steer conversations toward WME, even when prompted with AAL texts. Our findings provide a targeted analysis of anti-AAL biases at a relatively understudied stage in LLM development, highlighting representational harms and ethical questions about the desired behavior of LLMs concerning AAL.

Warning: This paper contains potentially offensive and harmful text. Large Reasoning Models (LRMs) have recently extended their powerful reasoning capabilities to safety checks--using chain-of-thought reasoning to decide whether a request should be answered. While this new approach offers a promising route for balancing model utility and safety, its robustness remains underexplored. To address this gap, we introduce Malicious-Educator, a benchmark that disguises extremely dangerous or malicious requests beneath seemingly legitimate educational prompts. Our experiments reveal severe security flaws in popular commercial-grade LRMs, including OpenAI o1/o3, DeepSeek-R1, and Gemini 2.0 Flash Thinking. For instance, although OpenAI's o1 model initially maintains a high refusal rate of about 98%, subsequent model updates significantly compromise its safety; and attackers can easily extract criminal strategies from DeepSeek-R1 and Gemini 2.0 Flash Thinking without any additional tricks. To further highlight these vulnerabilities, we propose Hijacking Chain-of-Thought (H-CoT), a universal and transferable attack method that leverages the model's own displayed intermediate reasoning to jailbreak its safety reasoning mechanism. Under H-CoT, refusal rates sharply decline--dropping from 98% to below 2%--and, in some instances, even transform initially cautious tones into ones that are willing to provide harmful content. We hope these findings underscore the urgent need for more robust safety mechanisms to preserve the benefits of advanced reasoning capabilities without compromising ethical standards.

With the power of LLMs, we now have the ability to query data that was previously impossible to query, including text, images, and video. However, despite this enormous potential, most present-day data systems that leverage LLMs are reactive, reflecting our community's desire to map LLMs to known abstractions. Most data systems treat LLMs as an opaque black box that operates on user inputs and data as is, optimizing them much like any other approximate, expensive UDFs, in conjunction with other relational operators. Such data systems do as they are told, but fail to understand and leverage what the LLM is being asked to do (i.e. the underlying operations, which may be error-prone), the data the LLM is operating on (e.g., long, complex documents), or what the user really needs. They don't take advantage of the characteristics of the operations and/or the data at hand, or ensure correctness of results when there are imprecisions and ambiguities. We argue that data systems instead need to be proactive: they need to be given more agency -- armed with the power of LLMs -- to understand and rework the user inputs and the data and to make decisions on how the operations and the data should be represented and processed. By allowing the data system to parse, rewrite, and decompose user inputs and data, or to interact with the user in ways that go beyond the standard single-shot query-result paradigm, the data system is able to address user needs more efficiently and effectively. These new capabilities lead to a rich design space where the data system takes more initiative: they are empowered to perform optimization based on the transformation operations, data characteristics, and user intent. We discuss various successful examples of how this framework has been and can be applied in real-world tasks, and present future directions for this ambitious research agenda.

Patent analysis highly relies on concise and interpretable document representations, referred to as patent portraits. Keyphrases, both present and absent, are ideal candidates for patent portraits due to their brevity, representativeness, and clarity. In this paper, we introduce KAPPA, an integrated framework designed to construct keyphrase-based patent portraits and enhance patent analysis. KAPPA operates in two phases: patent portrait construction and portrait-based analysis. To ensure effective portrait construction, we propose a semantic-calibrated keyphrase generation paradigm that integrates pre-trained language models with a prompt-based hierarchical decoding strategy to leverage the multi-level structural characteristics of patents. For portrait-based analysis, we develop a comprehensive framework that employs keyphrase-based patent portraits to enable efficient and accurate patent analysis. Extensive experiments on benchmark datasets of keyphrase generation, the proposed model achieves significant improvements compared to state-of-the-art baselines. Further experiments conducted on real-world patent applications demonstrate that our keyphrase-based portraits effectively capture domain-specific knowledge and enrich semantic representation for patent analysis tasks.

We present an ethical decision-making framework that refines a pre-trained reinforcement learning (RL) model using a task-agnostic ethical layer. Following initial training, the RL model undergoes ethical fine-tuning, where human feedback is replaced by feedback generated from a large language model (LLM). The LLM embodies consequentialist, deontological, virtue, social justice, and care ethics as moral principles to assign belief values to recommended actions during ethical decision-making. An ethical layer aggregates belief scores from multiple LLM-derived moral perspectives using Belief Jensen-Shannon Divergence and Dempster-Shafer Theory into probability scores that also serve as the shaping reward, steering the agent toward choices that align with a balanced ethical framework. This integrated learning framework helps the RL agent navigate moral uncertainty in complex environments and enables it to make morally sound decisions across diverse tasks. Our approach, tested across different LLM variants and compared with other belief aggregation techniques, demonstrates improved consistency, adaptability, and reduced reliance on handcrafted ethical rewards. This method is especially effective in dynamic scenarios where ethical challenges arise unexpectedly, making it well-suited for real-world applications.