This technical report accompanies the LLM robustness leaderboard published by PRISM Eval for the Paris AI Action Summit. We introduce PRISM Eval Behavior Elicitation Tool (BET), an AI system performing automated red-teaming through Dynamic Adversarial Optimization that achieves 100% Attack Success Rate (ASR) against 37 of 41 state-of-the-art LLMs. Beyond binary success metrics, we propose a fine-grained robustness metric estimating the average number of attempts required to elicit harmful behaviors, revealing that attack difficulty varies by over 300-fold across models despite universal vulnerability. We introduce primitive-level vulnerability analysis to identify which jailbreaking techniques are most effective for specific hazard categories. Our collaborative evaluation with trusted third parties from the AI Safety Network demonstrates practical pathways for distributed robustness assessment across the community.

Recommendation systems often suffer from data sparsity, caused by limited user-item interactions, which degrades their performance and amplifies popularity bias in real-world scenarios. This paper proposes a novel data augmentation framework that leverages Large Language Models (LLMs) and item textual descriptions to enrich interaction data. By few-shot prompting LLMs multiple times to rerank items and aggregating the results via majority voting, we generate high-confidence synthetic user-item interactions, supported by theoretical guarantees based on the concentration of measure. To effectively leverage the augmented data in the context of a graph recommendation system, we integrate it into a graph contrastive learning framework to mitigate distributional shift and alleviate popularity bias. Extensive experiments show that our method improves accuracy and reduces popularity bias, outperforming strong baselines.

We introduce MapStory, an LLM-powered animation prototyping tool that generates editable map animation sequences directly from natural language text by leveraging a dual-agent LLM architecture. Given a user written script, MapStory automatically produces a scene breakdown, which decomposes the text into key map animation primitives such as camera movements, visual highlights, and animated elements. Our system includes a researcher agent that accurately queries geospatial information by leveraging an LLM with web search, enabling automatic extraction of relevant regions, paths, and coordinates while allowing users to edit and query for changes or additional information to refine the results. Additionally, users can fine-tune parameters of these primitive blocks through an interactive timeline editor. We detail the system's design and architecture, informed by formative interviews with professional animators and by an analysis of 200 existing map animation videos. Our evaluation, which includes expert interviews (N=5) and a usability study (N=12), demonstrates that MapStory enables users to create map animations with ease, facilitates faster iteration, encourages creative exploration, and lowers barriers to creating map-centric stories.

Large language models (LLMs) have become integral to various real-world applications, leveraging massive, web-sourced datasets like Common Crawl, C4, and FineWeb for pretraining. While these datasets provide linguistic data essential for high-quality natural language generation, they often contain harmful content, such as hate speech, misinformation, and biased narratives. Training LLMs on such unfiltered data risks perpetuating toxic behaviors, spreading misinformation, and amplifying societal biases which can undermine trust in LLM-driven applications and raise ethical concerns about their use. This paper presents a large-scale analysis of inappropriate content across these datasets, offering a comprehensive taxonomy that categorizes harmful webpages into Topical and Toxic based on their intent. We also introduce a prompt evaluation dataset, a high-accuracy Topical and Toxic Prompt (TTP), and a transformer-based model (HarmFormer) for harmful content filtering. Additionally, we create a new multi-harm open-ended toxicity benchmark (HA VOC) and provide crucial insights into how models respond to adversarial toxic inputs. Our work offers insights into ensuring safer LLM pretraining and serves as a resource for Responsible AI (RAI) compliance. Disclaimer: This paper includes potentially offensive content due to the nature of the research.

Recently, GPT-4o has garnered significant attention for its strong performance in image generation, yet open-source models still lag behind. Several studies have explored distilling image data from GPT-4o to enhance open-source models, achieving notable progress. However, a key question remains: given that real-world image datasets already constitute a natural source of high-quality data, why should we use GPT-4o-generated synthetic data? In this work, we identify two key advantages of synthetic images. First, they can complement rare scenarios in real-world datasets, such as surreal fantasy or multi-reference image generation, which frequently occur in user queries. Second, they provide clean and controllable supervision. Real-world data often contains complex background noise and inherent misalignment between text descriptions and image content, whereas synthetic images offer pure backgrounds and long-tailed supervision signals, facilitating more accurate text-to-image alignment. Building on these insights, we introduce Echo-4o-Image, a 180K-scale synthetic dataset generated by GPT-4o, harnessing the power of synthetic image data to address blind spots in real-world coverage. Using this dataset, we fine-tune the unified multimodal generation baseline Bagel to obtain Echo-4o. In addition, we propose two new evaluation benchmarks for a more accurate and challenging assessment of image generation capabilities: GenEval++, which increases instruction complexity to mitigate score saturation, and Imagine-Bench, which focuses on evaluating both the understanding and generation of imaginative content. Echo-4o demonstrates strong performance across standard benchmarks. Moreover, applying Echo-4o-Image to other foundation models (e.g., OmniGen2, BLIP3-o) yields consistent performance gains across multiple metrics, highlighting the datasets strong transferability.

ChatGPT's dangerous dietary recommendation led to a man replacing table salt with toxic sodium bromide for three months, causing severe poisoning and hospitalization.

Evaluating audio generation systems, including text-to-music (TTM), text-to-speech (TTS), and text-to-audio (TTA), remains challenging due to the subjective and multi-dimensional nature of human perception. Existing methods treat mean opinion score (MOS) prediction as a regression problem, but standard regression losses overlook the relativity of perceptual judgments. To address this limitation, we introduce QAMRO, a novel Quality-aware Adaptive Margin Ranking Optimization framework that seamlessly integrates regression objectives from different perspectives, aiming to highlight perceptual differences and prioritize accurate ratings. Our framework leverages pre-trained audio-text models such as CLAP and Audiobox-Aesthetics, and is trained exclusively on the official AudioMOS Challenge 2025 dataset. It demonstrates superior alignment with human evaluations across all dimensions, significantly outperforming robust baseline models.

We introduce DiffPhysCam, a differentiable camera simulator designed to support robotics and embodied AI applications by enabling gradient-based optimization in visual perception pipelines. Generating synthetic images that closely mimic those from real cameras is essential for training visual models and enabling end-to-end visuomotor learning. Moreover, differentiable rendering allows inverse reconstruction of real-world scenes as digital twins, facilitating simulation-based robotics training. However, existing virtual cameras offer limited control over intrinsic settings, poorly capture optical artifacts, and lack tunable calibration parameters -- hindering sim-to-real transfer. DiffPhysCam addresses these limitations through a multi-stage pipeline that provides fine-grained control over camera settings, models key optical effects such as defocus blur, and supports calibration with real-world data. It enables both forward rendering for image synthesis and inverse rendering for 3D scene reconstruction, including mesh and material texture optimization. We show that DiffPhysCam enhances robotic perception performance in synthetic image tasks. As an illustrative example, we create a digital twin of a real-world scene using inverse rendering, simulate it in a multi-physics environment, and demonstrate navigation of an autonomous ground vehicle using images generated by DiffPhysCam.

AI systems for music generation are increasingly common and easy to use, granting people without any musical background the ability to create music. Because of this, generative-AI has been marketed and celebrated as a means of democratizing music making. However, inclusivity often functions as marketable rhetoric rather than a genuine guiding principle in these industry settings. In this paper, we look at four generative-AI music making systems available to the public as of mid-2025 (AIVA, Stable Audio, Suno, and Udio) and track how they are rhetoricized by their developers, and received by users. Our aim is to investigate ideologies that are driving the early-stage development and adoption of generative-AI in music making, with a particular focus on democratization. A combination of autoethnography and digital ethnography is used to examine patterns and incongruities in rhetoric when positioned against product functionality. The results are then collated to develop a nuanced, contextual discussion. The shared ideology we map between producers and consumers is individualist, globalist, techno-liberal, and ethically evasive. It is a 'total ideology' which obfuscates individual responsibility, and through which the nature of music and musical practice is transfigured to suit generative outcomes.

Semantic text classification requires the understanding of the contextual significance of specific tokens rather than surface-level patterns or keywords (as in rule-based or statistical text classification), making large language models (LLMs) well-suited for this task. However, semantic classification applications in industry, like customer intent detection or semantic role labeling, tend to be highly specialized. They require annotation by domain experts in contrast to general-purpose corpora for pretraining. Further, they typically require high inference throughputs which limits the model size from latency and cost perspectives. Thus, for a range of specialized classification tasks, the preferred solution is to develop customized classifiers by finetuning smaller language models (e.g., mini-encoders, small language models). In this work, we develop a token-driven sparse finetuning strategy to adapt small language models to specialized classification tasks. We identify and finetune a small sensitive subset of model parameters by leveraging task-specific token constructs in the finetuning dataset, while leaving most of the pretrained weights unchanged. Unlike adapter approaches such as low rank adaptation (LoRA), we do not introduce additional parameters to the model. Our approach identifies highly relevant semantic tokens (case study in the Appendix) and outperforms end-to-end finetuning, LoRA, layer selection, and prefix tuning on five diverse semantic classification tasks. We achieve greater stability and half the training costs vs. end-to-end finetuning.