InEx: Hallucination Mitigation via Introspection and Cross-Modal Multi-Agent Collaboration

Recent advancements in multimodal large language models have led to the introduction of GeoViS, a Geospatially Rewarded Visual Search framework aimed at enhancing visual grounding in remote sensing imagery. This framework addresses the challenges of identifying small targets within expansive scenes by employing a progressive search-and-reasoning process that integrates multimodal perception and spatial reasoning.

A recent study introduces Multi-resolution Retrieval-Detection (MRD), a framework aimed at enhancing high-resolution image understanding by addressing the challenges faced by multimodal large language models (MLLMs) in processing fragmented image crops. This approach allows for better semantic similarity computation by handling objects of varying sizes at different resolutions.

A new study introduces a framework called UNIFIER, aimed at addressing catastrophic forgetting in Multimodal Large Language Models (MLLMs) during continual learning in visual understanding. The research constructs a multimodal visual understanding dataset (MSVQA) that includes diverse scenarios such as high altitude and underwater perspectives, enabling MLLMs to adapt effectively to dynamic visual tasks.

A new benchmark called PPTBench has been introduced to evaluate large language models (MLLMs) on PowerPoint-related tasks, addressing the gap in existing benchmarks that focus on narrow subtasks and neglect layout-centric challenges. PPTBench utilizes a diverse dataset of 958 PPTX files and assesses models across four categories: Detection, Understanding, Modification, and Generation, with a total of 4,439 samples.

A recent study has introduced a behavioral benchmark called BayesBench to evaluate the performance of large language models (LLMs) in multimodal integration tasks, inspired by psychophysics research. The study assesses nine LLMs, including GPT-5 Mini, through magnitude estimation tasks involving text and images, revealing insights into their implicit computational strategies and Bayesian behavior.

OmniBench has been introduced as a benchmark to evaluate the performance of omni-language models (OLMs) in processing visual, acoustic, and textual inputs simultaneously, highlighting the limitations of current open-source multimodal large language models (MLLMs) in instruction-following and reasoning tasks.

Researchers have found that large language models (LLMs) make decisions about networking and friendship in ways that closely resemble human behavior, both in synthetic simulations and real-world contexts. This suggests that LLMs can replicate social decision-making processes similar to those of people.

A new quantization method called Four Over Six (4/6) has been introduced to enhance the NVFP4 quantization algorithm, which is crucial for the efficient training and inference of large language models (LLMs). This method evaluates multiple scale factors for blocks of values, aiming to reduce quantization errors that can lead to performance degradation during model training and inference.