Brief

Researchers have introduced ReceiptBench, a new benchmark designed to evaluate Multimodal Large Language Models (MLLMs) on understanding real-world documents like receipts. The benchmark includes 10,000 diverse receipts and is structured into four hierarchical tasks, ranging from basic text spotting to complex structure parsing and semantic reasoning. To improve MLLM performance on these tasks, a novel two-stage training framework called Metric-Aware Group Relative Policy Optimization (GRPO) was developed, which uses evaluation metrics as reinforcement learning signals for enhanced structural consistency. AI

Researchers have developed CVSearch, a new framework designed to improve how multimodal large language models (MLLMs) process high-resolution images. This training-free system dynamically adapts its search strategy, first attempting an expert-assisted search and then employing a novel semantic-aware scanning mechanism if the initial attempt fails. CVSearch aims to overcome the efficiency and coverage trade-offs of existing methods by intelligently decomposing images and exploring details iteratively, achieving state-of-the-art accuracy while enhancing search efficiency. AI

Researchers have developed a new method called Adversarial Subspace Alignment (ASAM) to improve knowledge editing in multimodal large language models (MLLMs). This technique addresses the limitation of current methods that struggle to generalize edits across semantically similar visual and linguistic variations. ASAM introduces Latent Adversarial Robustification (LAR) to identify and exploit fragile semantic regions, and Rank-Constrained Subspace Learning (RCSL) to align representations and ensure consistent predictions within knowledge units. AI

Researchers are developing multimodal large language models (MLLMs) that can process and integrate information from various data types, including text, audio, and video. One approach, MM-When2Speak, focuses on improving conversational timing by predicting when a brief reaction or a full response is appropriate, showing a threefold improvement in performance. Other research explores training MLLMs using only pairwise modalities to reduce data curation effort and addresses fine-grained visual understanding challenges through self-distillation techniques. These advancements aim to create more natural, engaging, and capable AI systems that can better perceive and interact with the real world. AI

Researchers have developed FashionLens, a unified framework for versatile fashion image retrieval using Multimodal Large Language Models. This system addresses the limitations of existing approaches by supporting diverse query formats and search intentions. To achieve this, FashionLens incorporates a Proposal-Guided Spherical Query Calibrator for task-aligned metric spaces and a Gradient-Guided Adaptive Sampling strategy to balance optimization across varying task complexities. The framework demonstrates state-of-the-art performance on the new U-FIRE benchmark, which consolidates fragmented fashion datasets. AI

Researchers have introduced a new benchmark and dataset called MM-OCEAN to evaluate how well multimodal large language models (MLLMs) can reason about personality. The study found that a significant portion of MLLMs, over 51%, provide correct personality assessments without grounding their judgments in observable behavioral evidence. This "Prejudice Gap" highlights a disconnect between accurate predictions and genuine understanding, suggesting a need for more robust evaluation methods for social cognition in AI. AI

Researchers have introduced SVFSearch, a new benchmark designed to evaluate multimodal large language models in short-video frame search, specifically within the Chinese gaming domain. The benchmark includes 5,000 test examples and 4,198 training examples, featuring paused game scenes from real short-video clips. SVFSearch provides a controlled environment with a game-domain corpus and image gallery to ensure reproducible evaluations, revealing significant gaps between model performance and oracle knowledge, and highlighting issues in visual grounding and retrieval. AI

Researchers have developed Faithful-MR1, a new training framework designed to improve the faithfulness of multimodal reasoning in large language models. This framework addresses the challenge of accurately perceiving and utilizing visual information during reasoning by anchoring and reinforcing visual attention. Experiments show Faithful-MR1 outperforms existing baselines on Qwen2.5-VL-Instruct models with less training data. Separately, another paper critiques the trustworthiness of current Vision-Language Models, arguing they often rely on language priors rather than genuine visual understanding and proposing new metrics to evaluate this 'Expense of Seeing'. AI

Researchers have introduced several new benchmarks and methods for Visual Question Answering (VQA) systems. HyLoVQA proposes a dynamic hypernetwork-generated low-rank adaptation technique for continual VQA, improving adaptation to new tasks and objects. WikiVQABench offers a knowledge-grounded VQA benchmark using Wikipedia and Wikidata, designed to test models requiring external knowledge. Additionally, UCSF-PDGM-VQA focuses on brain tumor MRI interpretation, highlighting current VLM limitations in clinical settings, while RoboSurg-VQA addresses surgical segmentation-aware VQA, and VISTAQA benchmarks joint answer correctness and pixel-level evidence grounding. AI

Researchers have developed FRA-Attack, a novel method to improve the transferability of adversarial attacks against multimodal large language models (MLLMs). This technique utilizes frequency-domain regularization to align perturbations with shared visual cues across different models, overcoming limitations of existing spatial-domain approaches. Experiments on 15 MLLMs demonstrate FRA-Attack's superior performance, particularly against models like GPT-5.4, Claude-Opus-4.6, and Gemini-3-flash. AI

Researchers have developed new benchmarks to evaluate the spatial reasoning capabilities of vision-language models (VLMs). ArchSIBench focuses on architectural space understanding, while Flat-Pack Bench assesses spatio-temporal reasoning in tasks like furniture assembly. SpaceDG addresses robustness by evaluating models under visual degradation, finding that current VLMs struggle with these challenges. Additionally, a framework called SAGE aims to improve spatial reasoning by enforcing geometric logic consistency. AI

Researchers have introduced two new benchmarks, VGenST-Bench and CaST-Bench, designed to more rigorously evaluate the spatio-temporal reasoning capabilities of Multimodal Large Language Models (MLLMs) and Vision-Language Models (VLMs). VGenST-Bench utilizes active video synthesis to create controlled scenarios across various spatial and temporal dimensions, enabling fine-grained diagnosis of MLLM understanding. CaST-Bench focuses on causal chain-grounded spatio-temporal reasoning, requiring models to identify and localize evidence for cause-and-effect relationships in videos, highlighting current VLM limitations in this area. AI

Two new research papers highlight challenges in developing AI for non-English languages and cultures. One paper reflects on two decades of building Arabic NLP resources, concluding that social and institutional factors are harder to overcome than linguistic ones. The other paper introduces a benchmark for evaluating how well Multimodal Large Language Models (MLLMs) can adapt to different cultures without negatively impacting their performance in other cultural contexts. AI