Brief

Researchers have introduced TextSculptor, a new framework designed to improve scene text editing in images. This framework includes an automated data construction pipeline that generates a large dataset of 3.2 million samples for text-to-image synthesis and text editing tasks. Additionally, TextSculptor provides a benchmark suite covering four core editing functions: addition, replacement, removal, and hybrid editing, aiming to enhance the performance of open-source models in this domain. AI

Researchers have identified a temporal grounding issue in multimodal large language models (MLLMs) where the models understand event timing during an initial phase but lose this signal during answer generation. They discovered specific attention heads, termed Temporal Grounding Heads (TG-Heads), that focus on the correct time intervals in videos during prefill. To address this, they developed an inference-time framework that leverages these TG-Heads to extract the relevant interval and then re-invokes the model with restricted visual context, improving performance on video temporal grounding benchmarks without requiring model retraining. AI

Researchers have developed SkeletonLLM, a novel approach to enable multimodal large language models (MLLMs) to understand structured, non-visual data like human skeletons. The system uses DrAction, a differentiable renderer that converts skeletal motion into image sequences, allowing MLLMs to process this data directly. This method facilitates open-vocabulary action recognition, motion captioning, and question answering across diverse skeleton formats, suggesting a path for MLLMs to engage with non-native data types. AI

Researchers have developed ST-SimDiff, a novel framework designed to make multimodal large language models (MLLMs) more efficient at processing long videos. The method addresses the computational burden by focusing on both static redundancy and dynamic changes within video content. ST-SimDiff utilizes a spatio-temporal graph to model token associations, employing a dual-selection strategy that identifies representative tokens for static information and key turning points for dynamic content. Experiments indicate that this approach significantly outperforms existing methods while reducing computational costs. AI

Researchers have developed a new pipeline to improve the ability of multimodal large language models (MLLMs) to analyze safety-critical driving events. This pipeline fuses downsampled video frames with telematics data and insights from specialized computer vision models to create high-quality training data. By fine-tuning the open-source QwenVL-2.5 model using this data, they achieved significant improvements in identifying and explaining safety-critical events with a limited computational budget. AI

Researchers have introduced MapTab, a new benchmark designed to evaluate the multi-criteria reasoning abilities of multimodal large language models (MLLMs). This benchmark utilizes route planning tasks that combine visual map data with structured tabular information on criteria such as time and price. MapTab includes two scenarios, Metromap and Travelmap, featuring extensive datasets of maps, queries, and questions to challenge MLLMs. Initial evaluations indicate that current MLLMs struggle with these complex multimodal reasoning tasks, sometimes underperforming unimodal approaches when visual perception is limited. AI

Researchers have introduced BEAR, a new benchmark designed to evaluate and diagnose the skill-level capabilities of embodied multimodal large language models (MLLMs). This benchmark decomposes embodied tasks into 14 distinct atomic skills, providing more granular insights into model failures than previous task-level evaluations. Evaluations on BEAR revealed that perceptual limitations and unstable spatiotemporal modeling are significant bottlenecks for current MLLMs. To address these issues, the team developed BEAR-Agent, a conversational agent that enhances MLLMs with visual and spatial reasoning tools, demonstrating substantial performance improvements on the benchmark and in robotic experiments. AI

Researchers have introduced AtelierEval, a novel benchmark designed to evaluate the proficiency of both humans and multimodal large language models (MLLMs) in generating effective text-to-image prompts. This benchmark, which includes 360 expert-crafted tasks, aims to quantify the quality of prompts used to translate user intent into detailed instructions for text-to-image systems. AtelierEval also features AtelierJudge, an agentic evaluator that correlates strongly with human expert assessments, and its experiments reveal that mimicry-based prompting may be more effective than planning-based approaches for future prompters. AI

Researchers have developed a new framework called AKT-Rec to address challenges in long-tail recommendation systems, particularly those in e-commerce platforms with significant data imbalance. This framework utilizes multimodal large language models (MLLMs) to generate semantic IDs that align content features with collaborative filtering signals. AKT-Rec incorporates an asymmetric contrastive objective and an activity-aware gating mechanism to facilitate knowledge transfer from head to tail items, improving representation learning. Experiments on a large-scale industrial dataset and subsequent online A/B testing on Alibaba's Tmall platform demonstrated substantial improvements in key metrics such as AUC, GAUC, CTR, and GMV. AI

Researchers have introduced EgoCoT-Bench, a new benchmark designed to evaluate the reasoning capabilities of Multimodal Large Language Models (MLLMs) when processing egocentric video data. This benchmark specifically focuses on the models' ability to understand hand-object interactions, track object states, and reason about manipulative processes using first-person video perspectives. EgoCoT-Bench aims to address limitations in existing benchmarks by providing explicit, step-by-step rationale annotations grounded in spatio-temporal evidence, revealing that many current MLLMs generate correct answers with inconsistent supporting evidence. AI