SpatialReasoner: Active Perception for Large-Scale 3D Scene Understanding

A novel self-supervised symbolic auto-encoder has been introduced, enabling symbolic computer vision to interpret diagrams through structured representations and logical rules. This approach contrasts with traditional pixel-based visual models by parsing diagrams into geometric primitives, enhancing machine vision's interpretability.

A comparative study has been conducted on the object counting capabilities of two multi-modal large language models, GPT-4o and GPT-5, focusing on their performance in zero-shot scenarios using only textual prompts. The evaluation was carried out on the FSC-147 and CARPK datasets, revealing that both models achieved results comparable to state-of-the-art methods, with some instances exceeding them.

A new framework called 'Look, Recite, Then Answer' has been proposed to enhance the performance of Vision-Language Models (VLMs) by generating self-generated knowledge hints. This approach aims to address the limitations of VLMs in specialized fields like precision agriculture, where reasoning-driven hallucination can hinder accurate visual perception.

The introduction of DIQ-H marks a significant advancement in evaluating the robustness of Vision-Language Models (VLMs) under conditions of temporal visual degradation, addressing critical failure modes such as hallucination persistence. This benchmark applies various physics-based corruptions to assess how VLMs recover from errors across multiple frames in dynamic environments.

A new method for Scene Graph Anticipation (SGA) has been introduced, termed Linguistic Scene Graph Anticipation (LSGA), which utilizes a language-driven framework to enhance the prediction of future scene graphs from video clips. This approach aims to improve the understanding of dynamic scenes by integrating semantic dynamics and commonsense temporal regularities, which are often difficult to extract from visual features alone.

SPARK introduces a three-stage framework for reinforcement learning that utilizes process reward models (PRMs) to provide dense feedback without the need for costly annotations. The first stage involves generating diverse solutions, which are then evaluated by a verifier model, leading to the creation of synthetic training data for fine-tuning PRMs. This method has demonstrated superior performance on benchmarks like ProcessBench, achieving an F1 score of 67.5 compared to traditional methods.

A new dataset and benchmark named UnicEdit-10M has been introduced to address the performance gap between closed-source and open-source multimodal models in image editing. This dataset, comprising 10 million entries, utilizes a lightweight data pipeline and a dual-task expert model, Qwen-Verify, to enhance quality control and failure detection in editing tasks.

A new method called Contextual Image Attack (CIA) has been proposed to exploit vulnerabilities in Multimodal Large Language Models (MLLMs) by embedding harmful queries within benign visual contexts. This approach utilizes a multi-agent system and various visualization strategies to enhance the attack's effectiveness, achieving high toxicity scores against models like GPT-4o and Qwen2.5-VL-72B.