The paper presents methods to reduce gender bias in audio‑enabled large language models (AuLLMs) by using adversarial training to unlearn gender perception without harming performance on other tasks. It also introduces a distillation approach that retains gender awareness but aligns responses with a gender‑agnostic teacher unless explicitly prompted. Experiments on three AuLLMs show promising results for gender‑unlearning and partial success for the distillation technique.
By Adnan Al Ali, Jind\v{r}ich Libovick\'y, Anne Lauscher
The paper argues that machine unlearning, which seeks to erase specific information from large language models, is often evaluated only on a single benchmark distribution, leading to an incomplete assessment of privacy. The authors demonstrate that forgetting can differ significantly when the same information is queried in different languages and scripts across multilingual models, revealing a gap between behavioral forgetting and true inaccessibility. They call for broader evaluation that includes access‑path generalization and adversarial recoverability to ensure genuine privacy protection.
By Tyler Lizzo, Larry Heck
The study evaluates name-based bias in six local large language models using a fill-in-the-blank task with first names from U.S. Social Security and a race classification dataset. Across 300,000 responses, the models consistently favored names that were less stereotypically male, older, popular, and white, with gender bias being the strongest effect. The bias varied with sentence context but remained significant, underscoring the persistence of name bias in local LLMs.
By Kalyan Bhetwal, Ceren Mert, Ian Whitehouse
The paper introduces OmniVerifier-C1, a graded visual verifier that decomposes complex prompts or environmental rules into fine‑grained components and assigns continuous pointwise scores, moving beyond binary True/False judgments. Experiments show that this structured, interpretable output improves generalization to out‑of‑domain multimodal critic tasks and outperforms binary verifiers in text‑to‑image scenarios. In rule‑governed generative world modeling, the authors demonstrate that a pairwise verification approach yields more accurate critiques than pointwise scoring, with empirical validation on tasks such as Maze and Sudoku.
By Bowei Liu, Xinchen Zhang, Xuhuan Li, Tianlin Zhang, Kaian Jiang, Jingwei Liu, Congyang Zhao, Yifei Qian, Junhong Liu, Zhiheng Li, Yujiu Yang, Yang Cai, Ling Yang
MagiC is a benchmark for grounded multimodal cognition that evaluates vision‑language models on answer accuracy, step‑by‑step reasoning quality, grounding fidelity, and self‑correction. It contains about 5,500 weakly supervised QA examples generated from strong model outputs and 900 human‑curated examples with fine‑grained annotations, including rationales and bounding box groundings. The benchmark tests 15 models ranging from 7B to 70B parameters, introduces new metrics such as MagiScore and StepSense, and includes diagnostic settings to probe robustness against adversarial visual cues and introspective error correction.
By Chengfei Wu, Ronald Seoh, Bingxuan Li, Liqiang Zhang, Fengrong Han, Dan Goldwasser
The study identifies a specific attention head (L8H10) in Vision Transformers that reads annotation text on chest X‑ray images instead of lung pathology, affecting 62% of COVID‑19 images. Causal inpainting experiments confirm that the head’s behavior is driven by annotation content, and removing the head increases lung attention by 1.52×. This spurious annotation‑attending behavior is consistent across multiple ViT architectures and datasets, persists after full fine‑tuning, and is invisible to standard output‑level evaluation, highlighting the need for mechanistic auditing.
By Joshua Muhumuza, Cornelius Kalenzi
MBBQ (Multilingual Bias Benchmark for Question-answering) is a curated extension of the English BBQ dataset that includes Dutch, Spanish, and Turkish, designed to measure stereotypes across these languages while controlling for cultural differences and task accuracy. The dataset also provides a parallel control set to assess question-answering performance independently of bias. Experiments with several open-source and proprietary LLMs show that non‑English languages can exhibit greater bias than English, and cross‑lingual bias differences persist except in the most accurate models.
By Vera Neplenbroek, Arianna Bisazza, Raquel Fern\'andez
MIXAR is the first generative multilingual pixel‑based language model, pretrained on eight high‑resource languages across Latin and CJK scripts with up to 0.5 B parameters. It replaces unstable adversarial training with a GRPO‑based reinforcement‑learning stage, improving free‑generation accuracy by up to 2.3× on multilingual benchmarks. MIXAR generalizes better to unseen low‑resource languages than a tokenizer‑based model trained on the same corpus, matches tokenizer models on discriminative tasks, and resists orthographic attacks more gracefully.
By Chen Hu, Yintao Tai, Antonio Vergari, Frank Keller, Alessandro Suglia
arXiv:2610. 14438v2 Announce Type: replace Abstract: Large Language Models (LLMs) are increasingly being used to summarize and verify claims in critical sources that inform consequential decisions.
By Duaa Nakshbandi, Zhiyu An, Yiping Li, Sachin Giri, Wan Du
The Digital Typhoon Dataset V3 extends the longest typhoon satellite image collection to nearly 50 years, now including multispectral satellite data for multi‑channel modeling. The study benchmarks machine learning models across the Northern and Southern Hemispheres, finding that multi‑channel inputs do not consistently outperform single‑channel ones for pressure regression, with ResNet‑18, full‑resolution images, and Channel 1 delivering the best single‑channel results. Systematic bias analysis reveals that intensity regression underestimates strong typhoons, and imbalance‑aware weighting strategies fail to fully correct this bias.
By Asanobu Kitamoto, Victor Le Louarn, Noelia Falczuk, Ngoc Anh Tong, Anja Veleva
Flow-CLIP introduces a method to improve compositional image-text matching by refining global image-text alignment with image-conditioned changes in token–patch evidence. It combines Global Compatibility Calibration (GCC) for sentence-level similarity, Image-Conditioned Evidence Refinement (ICER) to incorporate visual patch context, and Cross-Image Margin Ranking (CIMR) to enhance ranking margins when images support correct descriptions. Experiments demonstrate that Flow-CLIP outperforms pretrained CLIP and other global alignment approaches on binding accuracy and reduces reliance on language priors.
By Minghong Wei, Enguang Zhu, Pu Cao, Lu Yang, Qing Song
SCOPE introduces a reinforcement learning approach for text‑to‑image prompt engineering that respects semantic spans such as objects, counts, attributes, and relations. By pooling policy log‑probability shifts within each semantic span, the method assigns a shared coefficient to token losses while using a single advantage per completion, eliminating the need for separate span rewards or extra image renders. Experiments on Qwen‑Image, SD3, and FLUX.1‑dev show average relative gains of 9.48% on GeneVal and 7.62% on T2I‑CompBench compared to vanilla GRPO, with ablation studies confirming the benefit of semantic boundaries and pooling.
By Yufei Jiao, Jinfa Huang, Zeyue Xue, Tianyu Xie, Yuexiao Ma, Rongrong Ji, Xiawu Zheng, Jiebo Luo
Line-Augmented cuVSLAM extends the CUDA-accelerated visual‑SLAM system cuVSLAM by incorporating line features to improve visual odometry in low‑texture environments such as warehouses and production facilities. Lines are represented using Plücker coordinates for transport and a minimal 4‑DOF orthonormal form for optimisation, and a Cauchy‑robust point‑to‑line residual is integrated into the per‑frame Gauss‑Newton pose solve and sliding‑window bundle adjustment. The adaptive weighting allows line features to take on more of the estimation burden when point features are scarce, and the extension adds residual blocks to existing solvers rather than creating a separate estimator.
By Dara Molloy, Terry Wu, Wonju Lee, Abhilash Nand Kumar, David Adley, Conor Ferguson, Hesam Rabeti
This survey reviews 490 papers on text‑to‑image synthesis (T2I) over the past decade, covering key model families such as GANs, autoregressive models, and diffusion models. It organizes recent research by problem areas—controllability, personalization, editing, preference alignment, efficiency, safety, and spatial consistency—rather than strictly chronologically. The paper also examines unified multimodal models, datasets, evaluation metrics, and discusses applications, challenges, and future research directions in T2I.
By Nonghai Zhang, Hao Tang
SceneReVis is a vision‑grounded, self‑reflective framework for 3D indoor scene synthesis that uses an iterative diagnose‑and‑act loop to detect and fix spatial conflicts such as collisions. It introduces SceneChain‑12k, a large dataset of causal construction trajectories generated via a reverse engineering pipeline, and employs a two‑stage training recipe that moves from supervised fine‑tuning to agentic reinforcement learning, turning the model into an active spatial planner. Experiments show that SceneReVis achieves state‑of‑the‑art performance in high‑fidelity generation and goal‑oriented optimization, with better transfer to underrepresented room categories and potential for out‑of‑distribution room types.
By Yang Zhao, Shizhao Sun, Meisheng Zhang, Yingdong Shi, Xubo Yang, Jiang Bian
The paper introduces SADGE, a quantitative similarity metric that predicts how well synthetic image datasets will perform on real-world computer vision tasks without requiring downstream model training. SADGE combines appearance and geometric similarity measures through a constrained bilinear interaction, outperforming appearance-only and geometry-only baselines across 79,000 image pairs in five synthetic-to-real benchmark families. The study demonstrates that the non‑linear interplay between appearance and geometry, rather than either alone, best predicts downstream performance in tasks such as object detection, semantic segmentation, and pose estimation.
By Patryk Bartkowiak, Dominik Michels, Soren Pirk, Wojtek Palubicki
StepSafe introduces a training framework called ASPO that focuses on step‑level timely intervention in agentic systems. It aligns decisions with their respective prefixes, rewards actions that fit the current stage, and uses adaptive feedback self‑distillation to credit reasoning. The resulting StepSafe guardrail achieves state‑of‑the‑art performance on five agent‑safety benchmarks, improving the safety‑utility Pareto curve with a Youden’s J of 53.0 for a 4B model.
By Siyi Gu, Kunhan Gao, Chao Peter Yang, Sukanta Ganguly, Rex Ying, Arman Cohan
The paper introduces a three‑level shortcut‑auditing framework to assess AI guard reliability in agent‑safety evaluations. It demonstrates that shortcut cues—such as shorter shell commands or exact‑tool lookups—can predict trajectory labels and influence guard decisions, leading to significant changes in measured performance when benchmark representations are altered. Experiments on LinuxArena, AgentDojo‑Traj, and AgentHarm‑Traj datasets reveal that guard accuracy can drop dramatically under domain holdouts and that editing safe‑row markers or tool aliases materially affects false‑positive rates.
"whyItMatters":"The study shows that benchmark design can directly alter guard decisions and evaluation outcomes, underscoring the need for careful evaluation protocols to ensure reliable safety assessments."
By Hasan Demirkiran, Jens Ernstberger
The paper proposes a hierarchical conditioning approach for joint Acoustic Scene Classification and Audio Tagging, predicting scene categories directly from intermediate event probabilities. Extensive experiments on three backbones and varied loss‑weighting settings show that this method matches or surpasses traditional multi‑head models while enabling synchronized dual‑task learning for certain beta values. An interpretability framework combining information bottleneck analysis, manifold geometry, and Formal Concept Analysis on TF‑IDF event representations demonstrates that hierarchical conditioning reduces conditional entropy, increases mutual information, and improves latent cluster separability, yielding a more structured and semantically faithful feature space.
By Claudia Montero-Ram\'irez, Alba Mart\'inez-Serrano, Carmen Pel\'aez-Moreno
PanoDiff-SR is a two‑stage pipeline that generates full‑size dental panoramic radiographs (PRs) at diagnostic resolution. First, a denoising diffusion implicit model creates a low‑resolution PR seed, which is then upscaled by a fine‑tuned hybrid attention transformer to a high‑resolution image. The method achieves realistic results, with a Fréchet inception distance of 40.5 and a clinician detection accuracy of 68.5%, closer to chance than previous GAN‑based approaches.
By Sanyam Jain, Bruna Neves de Freitas, Andreas Basse-OConnor, Alexandros Iosifidis, Ruben Pauwels