Abstract--Large Language Models (LLMs) are increasingly used as automated judges for evaluating text quality, code c or-rectness, and argument strength. However, these LLM-as-a-judge systems are vulnerable to adversarial attacks that can mani pulate their assessments. This paper investigates the vulnerabil ity of LLM-as-a-judge systems to prompt injection attacks, drawi ng insights from both academic literature and practical solutio ns from the "LLMs: Y ou Can't Please Them All" Kaggle competition. We present a comprehensive framework for developing and evaluating adversarial attacks against LLM judges, distinguis hing between content-author attacks and system-prompt attacks . Through rigorous statistical a nalysis (n=50 prompts per condition, bootstrap confidence interval s), we demonstrate that sophisticated attacks can achieve succes s rates of up to 73.8% against popular LLM judges, with Contextual Misdirection being the most effective method against Gemma models at 67.7%. We find that smaller models like Gemma-3-4B-Instruct are more vulnerable (65.9% average success rat e) than their larger counterparts, and that attacks show high transferability (50.5-62.6%) We compare our approach with recent work including Universal-Prompt-Injection [1] and AdvPrompter [2], demonstrating b oth complementary insights and novel contributions. Our findin gs highlight critical vulnerabilities in current LLM-as-a-j udge systems and provide recommendations for developing more robus t evaluation frameworks, including using multi-model commi ttees with diverse architectures and preferring comparative ass essment over absolute scoring methods. T o ensure reproducibility, we release our code, evaluation harness, and processed datase ts.

Using multiple agents was found to improve the debugging capabilities of Large Language Models. However, increasing the number of LLM-agents has several drawbacks such as increasing the running costs and rising the risk for the agents to lose focus. In this work, we propose an adaptive agentic design, where the number of agents and their roles are determined dynamically based on the characteristics of the task to be achieved. In this design, the agents roles are not predefined, but are generated after analyzing the problem to be solved. Our initial evaluation shows that, with the adaptive design, the number of agents that are generated depends on the complexity of the buggy code. In fact, for simple code with mere syntax issues, the problem was usually fixed using one agent only. However, for more complex problems, we noticed the creation of a higher number of agents. Regarding the effectiveness of the fix, we noticed an average improvement of 11% compared to the one-shot prompting. Given these promising results, we outline future research directions to improve our design for adaptive software agents that can autonomously plan and conduct their software goals.

Using basic health statements authorized by UK and EU registers and ~9,100 journalist - vetted public - health assertions on topics such as abortion, COVID - 19 and politics from sources ranging from peer - reviewed journals and government advisories to social med ia and news across the political spectrum, we benchmark six leading large language models from in 21 languages, finding that -- despite high accuracy on English - centric textbook claims -- performance falls in multiple non - European languages and fluctuates by top ic and source, highlighting the urgency of comprehensive multilingual, domain - aware validation before deploying AI in global health communication. Main Text: Recent evidence suggests that 17 % of U.S. adults -- and a striking 25 % of those aged 18 - 29 -- now consult AI chatbots for health questions at least once a month (1), while in Australia nearly 10 % of adults did so in just the first half of 2024 (2). Beyond mere curiosity, these tools can substantially improve comprehension: running standard d ischarge notes through GPT - 4 reduced the average reading grade level from 11th to 6th and boosted patient - understandability scores from 13 % to 81 % (3). Yet as fluently as large language models (LLMs) can rephrase medical text, they lack formal clinical v etting and still rely on statistical patterns in their training data. When generative AI echoes unverified or dangerous claims, it risks amplifying harm.

Recent audio-visual generative models have made substantial progress in generating images from audio. However, existing approaches focus on generating images from single-class audio and fail to generate images from mixed audio. To address this, we propose an Audio-Visual Generation and Separation model (AV-GAS) for generating images from soundscapes (mixed audio containing multiple classes). Our contribution is threefold: First, we propose a new challenge in the audio-visual generation task, which is to generate an image given a multi-class audio input, and we propose a method that solves this task using an audio-visual separator. Second, we introduce a new audio-visual separation task, which involves generating separate images for each class present in a mixed audio input. Lastly, we propose new evaluation metrics for the audio-visual generation task: Class Representation Score (CRS) and a modified R@K. Our model is trained and evaluated on the VGGSound dataset. We show that our method outperforms the state-of-the-art, achieving 7% higher CRS and 4% higher R@2* in generating plausible images with mixed audio.

Generating images from prompts containing specific entities requires models to retain as much entity-specific knowledge as possible. However, fully memorizing such knowledge is impractical due to the vast number of entities and their continuous emergence. To address this, we propose Text-based Intelligent Generation with Entity prompt Refinement (TextTIGER), which augments knowledge on entities included in the prompts and then summarizes the augmented descriptions using Large Language Models (LLMs) to mitigate performance degradation from longer inputs. To evaluate our method, we introduce WiT-Cub (WiT with Captions and Uncomplicated Background-explanations), a dataset comprising captions, images, and an entity list. Experiments on four image generation models and five LLMs show that TextTIGER improves image generation performance in standard metrics (IS, FID, and CLIPScore) compared to caption-only prompts. Additionally, multiple annotators' evaluation confirms that the summarized descriptions are more informative, validating LLMs' ability to generate concise yet rich descriptions. These findings demonstrate that refining prompts with augmented and summarized entity-related descriptions enhances image generation capabilities. The code and dataset will be available upon acceptance.

Automating structured clinical interviews could revolutionize mental healthcare accessibility, yet existing large language models (LLMs) approaches fail to align with psychiatric diagnostic protocols. We present MAGI, the first framework that transforms the gold-standard Mini International Neuropsychiatric Interview (MINI) into automatic computational workflows through coordinated multi-agent collaboration. MAGI dynamically navigates clinical logic via four specialized agents: 1) an interview tree guided navigation agent adhering to the MINI's branching structure, 2) an adaptive question agent blending diagnostic probing, explaining, and empathy, 3) a judgment agent validating whether the response from participants meet the node, and 4) a diagnosis Agent generating Psychometric Chain-of- Thought (PsyCoT) traces that explicitly map symptoms to clinical criteria. Experimental results on 1,002 real-world participants covering depression, generalized anxiety, social anxiety and suicide shows that MAGI advances LLM- assisted mental health assessment by combining clinical rigor, conversational adaptability, and explainable reasoning.

We introduce a new generation of small reasoning models for RAG, search, and source summarization. Pleias-RAG-350m and Pleias-RAG-1B are mid-trained on a large synthetic dataset emulating the retrieval of a wide variety of multilingual open sources from the Common Corpus. They provide native support for citation and grounding with literal quotes and reintegrate multiple features associated with RAG workflows, such as query routing, query reformulation, and source reranking. Pleias-RAG-350m and Pleias-RAG-1B outperform SLMs below 4 billion parameters on standardized RAG benchmarks (HotPotQA, 2wiki) and are competitive with popular larger models, including Qwen-2.5-7B, Llama-3.1-8B, and Gemma-3-4B. They are the only SLMs to date maintaining consistent RAG performance across leading European languages and ensuring systematic reference grounding for statements. Due to their size and ease of deployment on constrained infrastructure and higher factuality by design, the models unlock a range of new use cases for generative AI.

This study examines the variability of Chat-GPT machine translation (MT) outputs across six different configurations in four languages,with a focus on creativity in a literary text. We evaluate GPT translations in different text granularity levels, temperature settings and prompting strategies with a Creativity Score formula. We found that prompting ChatGPT with a minimal instruction yields the best creative translations, with "Translate the following text into [TG] creatively" at the temperature of 1.0 outperforming other configurations and DeepL in Spanish, Dutch, and Chinese. Nonetheless, ChatGPT consistently underperforms compared to human translation (HT).

The integration of language models in the legal domain holds considerable promise for streamlining processes and improving efficiency in managing extensive workloads. However, the specialized terminology, nuanced language, and formal style of legal texts can present substantial challenges. This study examines whether preference-based training techniques, specifically Reinforcement Learning from Human Feedback and Direct Preference Optimization, can enhance models' performance in generating Icelandic legal summaries that align with domain-specific language standards and user preferences. We compare models fine-tuned with preference training to those using conventional supervised learning. Results indicate that preference training improves the legal accuracy of generated summaries over standard fine-tuning but does not significantly enhance the overall quality of Icelandic language usage. Discrepancies between automated metrics and human evaluations further underscore the importance of qualitative assessment in developing language models for the legal domain.

Large Language Models (LLM) are typically trained on vast amounts of data from various sources. Even when designed modularly (e.g., Mixture-of-Experts), LLMs can leak privacy on their sources. Conversely, training such models in isolation arguably prohibits generalization. Large Language Models have brought much disruption in the field of Artificial Intelligence and have transformed various use-cases, from intelligent assistants (Dong et al., 2023) and code copilots (Chen et al., 2021) to agentic web browsing (Zheng et al., 2024) and enhanced tutoring (Ko-talwar et al., 2024). They have shown great scaling potential, devouring terabytes of raw textual or multi-modal data (Kaplan et al., 2020) without their performance plateauing. As this trend continues, all public resources will eventually be consumed. Therefore, tapping into private data silos will become the next significant source of information (Shumailov et al., 2024; Iacob et al., 2024). This introduces the need to orchestrate model training that is somehow separated per region or source. Maintaining separate models, though, quickly becomes intractable and burdensome. Private organizations can own data they want to use for their custom LLM but not expose it publicly Carlini et al. (2021); OpenAI (2023). For instance, client institutions may wish to train domain-specific Copilots (GitHub, 2024) without leaking proprietary information (Niu et al., 2023) to the public domain. To approach this problem, we draw from Modular Learning (Pfeiffer et al., 2023) for routing knowledge across parts of a neural network and adaptively serve to different domains. While off-the-shelf Mixture-of-Experts (MoE) models (Cai et al., 2024) adopt an architecture where different domains can share common parameters - thus enabling knowledge transfer. However, they can introduce privacy risks (Carlini et al., 2019) exactly because of this sharing. In addition, training an entire MoE model under Differential Privacy (DP) significantly reduces its utility as training a large shared backbone network over multiple domains requires adding large amounts of DP noise.