Benchmarks are pivotal in driving AI progress, and invalid benchmark questions frequently undermine their reliability. Manually identifying and correcting errors among thousands of benchmark questions is not only infeasible but also a critical bottleneck for reliable evaluation. In this work, we introduce a framework for systematic benchmark revision that leverages statistical analysis of response patterns to flag potentially invalid questions for further expert review. Our approach builds on a core assumption commonly used in AI evaluations that the mean score sufficiently summarizes model performance. This implies a unidimensional latent construct underlying the measurement experiment, yielding expected ranges for various statistics for each item. When empirically estimated values for these statistics fall outside the expected range for an item, the item is more likely to be problematic. Across nine widely used benchmarks, our method guides expert review to identify problematic questions with up to 84\% precision. In addition, we introduce an LLM-judge first pass to review questions, further reducing human effort. Together, these components provide an efficient and scalable framework for systematic benchmark revision.

Cognitive BASIC is a minimal, BASIC-style prompting language and in-model interpreter that structures large language model (LLM) reasoning into explicit, stepwise execution traces. Inspired by the simplicity of retro BASIC, we repurpose numbered lines and simple commands as an interpretable cognitive control layer. Modern LLMs can reliably simulate such short programs, enabling transparent multi-step reasoning inside the model. A natural-language interpreter file specifies command semantics, memory updates, and logging behavior. Our mental-model interpreter extracts declarative and procedural knowledge, detects contradictions, and produces resolutions when necessary. A comparison across three LLMs on a benchmark of knowledge extraction, conflict detection, and reasoning tasks shows that all models can execute Cognitive BASIC programs, with overall strong but not uniform performance.

Evaluating and measuring AI Safety Level (ASL) threats are crucial for guiding stakeholders to implement safeguards that keep risks within acceptable limits. ASL-3+ models present a unique risk in their ability to uplift novice non-state actors, especially in the realm of biosecurity. Existing evaluation metrics, such as LAB-Bench, BioLP-bench, and WMDP, can reliably assess model uplift and domain knowledge. However, metrics that better contextualize "real-world risks" are needed to inform the safety case for LLMs, along with scalable, open-ended metrics to keep pace with their rapid advancements. To address both gaps, we introduce MOCET, an interpretable and doubly-scalable metric (automatable and open-ended) that can quantify real-world risks.

While recent research suggests Large Language Models match human creative performance in divergent thinking tasks, visual creativity remains underexplored. This study compared image generation in human participants (Visual Artists and Non Artists) and using an image generation AI model (two prompting conditions with varying human input: high for Human Inspired, low for Self Guided). Human raters (N=255) and GPT4o evaluated the creativity of the resulting images. We found a clear creativity gradient, with Visual Artists being the most creative, followed by Non Artists, then Human Inspired generative AI, and finally Self Guided generative AI. Increased human guidance strongly improved GenAI's creative output, bringing its productions close to those of Non Artists. Notably, human and AI raters also showed vastly different creativity judgment patterns. These results suggest that, in contrast to language centered tasks, GenAI models may face unique challenges in visual domains, where creativity depends on perceptual nuance and contextual sensitivity, distinctly human capacities that may not be readily transferable from language models.

3D meshes are a critical building block for applications ranging from industrial design and gaming to simulation and robotics. Traditionally, meshes are crafted manually by artists, a process that is time-intensive and difficult to scale. To automate and accelerate this asset creation, autoregressive models have emerged as a powerful paradigm for artistic mesh generation. However, current methods to enhance quality typically rely on larger models or longer sequences that result in longer generation time, and their inherent sequential nature imposes a severe quality-speed trade-off. This sequential dependency also significantly complicates incremental editing. To overcome these limitations, we propose Mesh RAG, a novel, training-free, plug-and-play framework for autoregressive mesh generation models. Inspired by RAG for language models, our approach augments the generation process by leveraging point cloud segmentation, spatial transformation, and point cloud registration to retrieve, generate, and integrate mesh components. This retrieval-based approach decouples generation from its strict sequential dependency, facilitating efficient and parallelizable inference. We demonstrate the wide applicability of Mesh RAG across various foundational autoregressive mesh generation models, showing it significantly enhances mesh quality, accelerates generation speed compared to sequential part prediction, and enables incremental editing, all without model retraining.

This paper presents JGU Mainz's winning system for the BLP-2025 Shared Task on Code Generation from Bangla Instructions. We propose a multi-agent-based pipeline. First, a code-generation agent produces an initial solution from the input instruction. The candidate program is then executed against the provided unit tests (pytest-style, assert-based). Only the failing cases are forwarded to a debugger agent, which reruns the tests, extracts error traces, and, conditioning on the error messages, the current program, and the relevant test cases, generates a revised solution. Using this approach, our submission achieved first place in the shared task with a $Pass@1$ score of 95.4. We also make our code public.

Multimodal large language models suffer from substantial inference overhead since multimodal KV Cache grows proportionally with the visual input length. Existing multimodal KV Cache compression methods mostly rely on attention score to reduce cache size, which makes them are incompatible with established efficient attention kernels (e.g., FlashAttention) and ignores the contribution of value vectors to the attention output. In this work, we revisit multimodal KV Cache compression from the perspective of the KV matrices' distribution. First, we observe that frequency-domain energy of multimodal KV matrices is predominantly concentrated in low-frequency and extract this principal energy via a low-pass filter. Further, we find that removing KV pairs that deviate substantially from this principal energy leads to a pronounced performance drop, which we define as Outlier KVs. Considering Outlier KVs are more likely to encode features critical for inference, we propose FlashCache, a frequency-domain-guided, Outlier-KV-aware KV Cache compression framework. First, we introduce an Outlier KV Recognition Module that models the principal component of multimodal KV matrices in the frequency domain and preferentially retains KV pairs that significantly deviate from it. Furthermore, Dynamic Budget Allocation Module is designed to adaptively determine the per-layer KV Cache size to retain more Outlier KVs. Experiments on multiple MLLMs and benchmarks demonstrate that FlashCache outperforms state-of-the-art multimoal KV compression methods, achieving up to 1.69 times faster decoding with 80% lower KV memory usage while maintaining task performance.

This paper introduces Generative Augmented Reality (GAR) as a next-generation paradigm that reframes augmentation as a process of world re-synthesis rather than world composition by a conventional AR engine. GAR replaces the conventional AR engine's multi-stage modules with a unified generative backbone, where environmental sensing, virtual content, and interaction signals are jointly encoded as conditioning inputs for continuous video generation. We formalize the computational correspondence between AR and GAR, survey the technical foundations that make real-time generative augmentation feasible, and outline prospective applications that leverage its unified inference model. We envision GAR as a future AR paradigm that delivers high-fidelity experiences in terms of realism, interactivity, and immersion, while eliciting new research challenges on technologies, content ecosystems, and the ethical and societal implications.

Graphs provide a unified representation of semantic content and relational structure, making them a natural fit for domains such as molecular modeling, citation networks, and social graphs. Meanwhile, large language models (LLMs) have excelled at understanding natural language and integrating cross-modal signals, sparking interest in their potential for graph reasoning. Recent work has explored this by either designing template-based graph templates or using graph neural networks (GNNs) to encode structural information. In this study, we investigate how different strategies for encoding graph structure affect LLM performance on text-attributed graphs. Surprisingly, our systematic experiments reveal that: (i) LLMs leveraging only node textual descriptions already achieve strong performance across tasks; and (ii) most structural encoding strategies offer marginal or even negative gains. We show that explicit structural priors are often unnecessary and, in some cases, counterproductive when powerful language models are involved. This represents a significant departure from traditional graph learning paradigms and highlights the need to rethink how structure should be represented and utilized in the LLM era. Our study is to systematically challenge the foundational assumption that structure is inherently beneficial for LLM-based graph reasoning, opening the door to new, semantics-driven approaches for graph learning.

Audio-language pretraining holds promise for general-purpose audio understanding, yet remains underexplored compared to its vision counterpart. While vision-language models like CLIP serve as widely adopted foundations, existing audio-language models primarily excel at retrieval tasks with limited adoption as general-purpose encoders. We identify three key barriers: limited large-scale audio-text corpora, insufficient caption diversity, and lack of systematic exploration and evaluation. To this end, we introduce CaptionStew, a 10.7M caption dataset aggregating diverse open-source audio-text corpora across multiple domains and captioning styles. Using this resource, we conduct the first comprehensive evaluation comparing contrastive and captioning objectives for audio representation learning across speech, music, and environmental sound tasks. Our results demonstrate that audio-language pretraining yields competitive, transferable representations. Through systematic data-scaling experiments, we reveal complementary objective strengths: contrastive learning achieves superior data efficiency at smaller scales, while captioning demonstrates better scalability on language-involved audio understanding tasks. We also find that common supervised initialization practices provide diminishing returns at scale, challenging current approaches. These findings establish audio-language pretraining as a viable pathway toward general-purpose audio representations, guiding future research. To accelerate progress, we release data preparation recipes, training protocols, and pretrained models, paving the way toward universal audio understanding. Early advances relied on supervised learning, where models trained on labeled corpora were adapted to related downstream tasks or transferred across domains (Kong et al., 2020; Chen et al., 2022a; Snyder et al., 2018; Desplanques et al., 2020).