The paper investigates hallucination detection in black‑box large language models by leveraging two accessible signals: semantic entropy, which captures disagreement among sampled response meanings, and token‑level uncertainty derived from log‑probabilities. It introduces a TopK aggregation technique, a hybrid CoCoA method combining uncertainty with semantic dissimilarity, and two supervised approaches—Gated and Stacked—that integrate token and semantic features. Across seven benchmarks and four language models, the supervised Stacked method performs best in many cases, while TopK and CoCoA remain competitive without labeled data, though all methods require careful threshold calibration.
By Urja Pawar, Rajitha Ramanayake, Owen O'Neill, Nabeel Kemal, Abhishek Mandal, Houssem Chatbri, Christopher Martin
The study presents a multimodal BrainAGE framework that fuses structural T1-weighted MRI with DeepCBV maps—vascular information synthesized from non‑contrast MRI—to estimate brain age more accurately. Using two separate 3D convolutional neural networks, the combined model achieved a mean absolute error of 3.95 years on cognitively normal controls, outperforming single‑modality models. Saliency analyses showed MRI highlighted white matter and cortical atrophy, while DeepCBV emphasized vascular‑rich regions, and the integrated approach revealed stronger differentiation between stable and progressive MCI, indicating sensitivity to early vascular changes.
By Jordan Jomsky, Zongyu Li, Kay C. Igwe, Yiren Zhang, Max Lashley, Tal Nuriel, Andrew Laine, Scott A. Small, Jia Guo, for the Frontotemporal Lobar Degeneration Neuroimaging Initiative, for the Alzheimer's Disease Neuroimaging Initiative
The paper introduces SEAL, a training-time, parameter‑efficient defense that attaches a plug‑and‑play adapter to the shared expert component of Mixture‑of‑Experts models, and SEAL++, which adds an orthogonal constraint to preserve existing safety subspaces. By leveraging the always‑activated shared expert, SEAL mitigates the structural vulnerability of sparse routing to adversarial manipulation, reducing attack success rates by up to 60% with minimal impact on model capability. The approach is evaluated across six attack scenarios involving harmful prompting, jailbreaks, malicious fine‑tuning, and neuron pruning.
By Qingyu Meng, Yiwei Zha, Jiahuan Pei, Koen Hindriks, Herbert Bos, Min Chen
The paper introduces Temporal Context Routing (TCR), a method that aligns script timing with the shared temporal axis used for audio and video generation. By routing prompt guidance to specific temporal positions in both modalities, TCR dramatically improves shot boundary accuracy and dialogue timing on a set of test scripts, while preserving visual quality and audio‑visual sync. A user study confirms that participants prefer TCR across multiple evaluation dimensions.
By Yichen Liu, Quanwei Zhang, Haozhe Wang, Donghao Zhou, Xiaojie Li, Yang Shi, Jiaming Liu, Ruihua Huang, Yingtian Zou, Daquan Zhou
Multimodal Language Models as Text-to-Image Model Evaluators presents MT2IE, a framework where a multimodal large language model generates evaluation prompts and scores images, achieving higher correlation with human judgment than prior metrics. MT2IE recovers official T2I model rankings using only 20 prompts—far fewer than traditional benchmarks—and adapts prompts to each model’s performance, maintaining informative scoring ranges. The approach demonstrates that dynamic, interactive evaluation can replace static benchmarks as T2I models improve.
By Jiahui Chen, Candace Ross, Reyhane Askari-Hemmat, Koustuv Sinha, Melissa Hall, Amy Zhang, Michal Drozdzal, Adriana Romero-Soriano
DocHop is a new benchmark that tests multimodal large language models on integrated chart‑context reasoning within document‑style images. The benchmark presents narrative text that imposes multi‑step compositional constraints, while charts supply the data needed to answer questions grounded in semantic reference labels. It contains 2,074 examples across six task categories, generated via a stochastic logic‑first pipeline that controls reasoning depth and visual density, and shows a large performance gap between humans (over 90% accuracy) and the best models (62.83%).
By Zhuoran Yu, Le Thien Phuc Nguyen, Jaden Park, Xinyi Gu, Zexue He, Soochahn Lee, Rogerio Feris, Yong Jae Lee
FemWear is a parameter‑efficient wearable foundation model specifically tailored for women's health. It repurposes a pretrained multimodal wearable backbone by training only 239,236 encoder parameters—just 1.11% of the original 21.54M—using low‑rank residual adapters and causal task‑family heads to create a shared longitudinal representation for menstrual, symptom, affective, sleep/recovery, autonomic, activity, and pregnancy outcomes. Evaluations across six cohorts and 63 metrics show improvements in cycle‑phase macro‑F1 and reductions in mean absolute error for cramps, mood symptoms, and sleep problems, while maintaining the OpenMHC ability‑retention benchmark.
By Yifan Wang, Chenzhong Li
Vision‑Language Models (VLMs) are increasingly replacing traditional OCR for document understanding, but this study shows they often rewrite imperfect text into more plausible forms, a flaw that clean‑text OCR benchmarks miss. The authors created FaithC4, a multilingual perturbation benchmark of 1,455 single‑page documents with scramble, random substitution, and visually similar substitution attacks, and evaluated 15 systems across general‑purpose VLMs, OCR‑specialized VLMs, and traditional OCR pipelines. Results reveal that general‑purpose VLMs suffer up to 6.9 WER points under perturbation, OCR‑specialized VLMs 0.1–3.4 points, and traditional OCR less than 0.8 points on English; probing Qwen3‑VL‑4B shows rewriting occurs only when a perturbed word’s final‑layer representation remains close to the original, with short words (4–6 characters) rewritten up to 10% of the time.
whyItMatters":"The findings highlight a critical limitation of VLMs in document transcription, underscoring the need for robust evaluation benchmarks that capture rewriting behavior beyond clean‑text accuracy."
By Gwang Gook Lee, Kenan Emir Ak, Jay Mohta, Yan Xu, Dimitrios Dimitriadis
The paper introduces a new task called video object segmentation‑aware audio generation, which conditions sound synthesis on object‑level segmentation maps. It presents SAGANet, a multimodal generative model that uses visual segmentation masks, video, and textual cues to produce controllable audio for musical instruments, offering fine‑grained, visually localized control. The authors also release the Segmented Music Solos dataset of instrument performance videos with segmentation information to support this task and demonstrate that SAGANet outperforms current state‑of‑the‑art methods in controllable, high‑fidelity Foley synthesis.
By Ilpo Viertola, Vladimir Iashin, Esa Rahtu
The paper presents CycleGRPO, a reinforcement learning framework that unifies region understanding and localization for multimodal large language models (MLLMs). By treating the MLLM as both actor and critic, the method generates region captions and immediately grounds them back into spatial coordinates, using a token‑level cycle‑consistency reward that obviates the need for textual ground truths. Experiments on SAMTok demonstrate that CycleGRPO can bootstrap region captioning, VQA, grounded dialogue, and referring segmentation simultaneously, achieving consistent performance gains without task‑specific fine‑tuning.
By Xin Zhang, Haochen Wang, Yikang Zhou, Zhuochen Wang, Xiangtai Li, Robby T. Tan
OptimusKG is a multimodal biomedical labeled property graph that integrates structured and semi‑structured resources to preserve detailed, type‑specific metadata across molecular, anatomical, clinical, and environmental domains. The graph contains nearly 191,000 nodes, over 21.8 million edges, and more than 67 million property instances derived from 18 ontologies, with a top‑level schema that enforces node and edge constraints while retaining granular provenance. Validation using the PaperQA3 agent found that 70.0% of sampled edges are supported by literature evidence, and the graph offers a standardized resource for machine learning, knowledge‑grounded retrieval, and hypothesis generation in biomedical research.
By Lucas Vittor, Ayush Noori, I\~naki Arango, Joaqu\'in Polonuer, Sam Rodriques, Andrew White, David A. Clifton, Marinka Zitnik
The paper investigates how knowledge distillation from Vision Transformers to smaller CNNs can cause dimensional collapse in the student’s representation space. Using SVD and Shannon entropy, the authors show that cosine‑based distillation leads to a drastic reduction in effective rank, while adding an InfoNCE objective can double the rank but harms downstream accuracy due to signal dilution. They further demonstrate that a label‑aware contrastive objective (Supervised Contrastive distillation) can maintain or improve accuracy without unnecessary rank expansion, indicating that effective rank alone is not a reliable indicator of representation quality.
By Kabir Thayani
The paper introduces TIC‑Bench, a new benchmark for evaluating multimodal large language models on deeply interleaved text‑image contexts. It covers logical, temporal, and spatial association tasks, totaling 2,280 questions across eight specific types. The authors benchmarked ten state‑of‑the‑art MLLMs, finding a significant performance gap versus human experts and highlighting persistent challenges in integrating evidence across interleaved visual and textual inputs.
By Zihao Wang, Xi Xiang, Yuwen Sun, Yingyu Li, Yabo Zhang, Yihan Zeng, Fan Li, Wangmeng Zuo
The paper investigates why post‑training boosts reasoning more than perception in vision‑language models. Using a diagnostic framework with synthetic tasks, it finds a perception‑reasoning asymmetry: supervised fine‑tuning suffers from token imbalance, while reinforcement learning suffers from reward coupling. The authors propose reweighting losses and perception‑aware rewards, achieving up to 18.2‑point and 6.0‑point gains respectively, and show that these methods also improve real‑world visual reasoning by up to 3.3 points.
By Xueqing Wu, Yu-Chi Lin, Kai-Wei Chang, Nanyun Peng
Beauty assessments from Multimodal Large Language Models (MLLMs) are increasingly popular, prompting a study comparing 2,513 human ratings to four commercial AI models—Claude, Gemini, GPT, and Grok. The study found that MLLMs consistently rate faces more favorably and with a narrower range than humans, yet they maintain strong correlations with human judgments and accurately track the rank‑ordering of faces. While all models agree strongly with each other, Grok showed the lowest agreement with human ratings, and only face age emerged as a common predictor of attractiveness across humans and MLLMs.
By Santiago Grandas, Juan Sebastian Cely-Acosta, Mohit Mendiratta, Shafee Hassan, Macken Murphy
The paper investigates how vision‑language models (VLMs) express visual content in discourse‑appropriate forms, focusing on information structure (IS) in Hungarian where Topic and Focus occupy distinct syntactic positions. By comparing six VLMs to human participants in visually grounded question answering, the authors find that VLMs produce IS‑relevant constructions but over‑regularise, collapsing onto narrow response templates, whereas humans use variable strategies influenced by discourse status, grammatical role, and definiteness. The study argues that VLM evaluation should extend beyond content accuracy to consider how content is packaged for discourse.
By Marcell Fekete, Johannes Bjerva, Tam\'as K\'aldi
ChartAttack is a framework that evaluates how multimodal large language models (MLLMs) can be misled by design misleaders to produce charts that cause incorrect interpretations. The authors also present AttackViz, a chart question‑answering dataset that labels effective misleaders and their induced wrong answers. Experiments show that ChartAttack can reduce MLLM QA accuracy by 17.2 points in‑domain and 11.9 points cross‑domain, and that fine‑tuning on AttackViz improves robustness to misleading charts.
By Jesus-German Ortiz-Barajas, Jonathan Tonglet, Vivek Gupta, Iryna Gurevych
The paper introduces Cue2Narrate, a two‑stage pipeline that jointly predicts what visual events to narrate and when to insert the narration in long, untrimmed movie clips. It uses a dual‑head audio‑visual localizer to identify visual cue and narration windows, followed by a LoRA‑adapted vision‑language model that generates concise audio descriptions, trained with a Description Ranking Loss. The authors also present the LongLSMDC benchmark, comprising up to 8‑minute clips, and show that Cue2Narrate outperforms video‑only and audio‑only baselines by 5–12 points in average mAP and improves AD generation over fine‑tuned base VLMs.
By Akshita Gupta, Aditya Arora, Federico Tombari, Marcus Rohrbach, Anna Rohrbach
The paper introduces a black-box membership inference attack framework tailored for fine-tuned text-to-speech models, addressing challenges in query generation and representation engineering. It evaluates five query types, finding recitation queries most effective, and uses multi-level speech embeddings with temporal alignment for fine-grained comparison. Experiments on CosyVoice2, F5-TTS, and XTTS-v2 trained on VCTK and British Dialect datasets show high privacy leakage, with speaker-level AUC above 0.80 and record-level AUC between 0.80 and 0.90.
By Kunlin Cai, Kaiyuan Zhang, Zihang Xiang, Jinghuai Zhang, Abeer Alwan, Fnu Suya, Yuan Tian
The paper introduces Zone of Proximal Policy Optimization (ZPPO), a method that keeps a teacher model inside prompts rather than in the policy gradient to improve knowledge distillation for small students. ZPPO creates two types of reformulated prompts—Binary Candidate-included Questions (BCQ) and Negative Candidate-included Questions (NCQ)—to expose students to correct and incorrect responses, and uses a replay buffer to focus training on hard questions until the student’s accuracy improves. Experiments on the Qwen3.5 family with a 27B teacher across 31 benchmarks show that ZPPO outperforms both off‑policy and on‑policy distillation methods, especially at the smallest student scales.
By Byung-Kwan Lee, Ximing Lu, Shizhe Diao, Minki Kang, Saurav Muralidharan, Karan Sapra, Andrew Tao, Pavlo Molchanov, Yejin Choi, Yu-Chiang Frank Wang, Ryo Hachiuma