Evaluating Multimodal Large Language Models on Vertically Written Japanese Text

A recent study has explored the integration of visual and textual information in Multimodal Large Language Models (MLLMs), revealing that visual-text fusion occurs at specific layers within these models rather than uniformly across the network. The research highlights a late-stage

A recent study has introduced a framework aimed at mitigating hallucination issues in Multimodal Large Language Models (MLLMs) during Reinforcement Learning (RL) optimization. The research identifies key factors contributing to hallucinations, including over-reliance on visual reasoning and insufficient exploration diversity. The proposed framework incorporates modules for caption feedback, diversity-aware sampling, and conflict regularization to enhance model reliability.

A new model named GraphFusionSBR has been introduced to enhance session-based recommendation systems by effectively capturing implicit user intents while addressing issues like item interaction dominance and noisy sessions. This model integrates multiple channels, including knowledge graphs and hypergraphs, to improve recommendation accuracy across various domains such as e-commerce and multimedia.

A recent study introduced OrderProbe, a deterministic benchmark designed to evaluate the structural reconstruction capabilities of large language models (LLMs) using fixed four-character expressions in Chinese, Japanese, and Korean. This benchmark aims to address the challenges of sentence-level restoration from scrambled inputs, which often lack a unique solution.

A recent study has investigated the dynamics of Large Language Model (LLM) agent reviewers within an Elo-ranked review system, utilizing real-world conference paper submissions. The research involved multiple LLM reviewers with distinct personas engaging in multi-round review interactions, moderated by an Area Chair, and highlighted the impact of Elo ratings and reviewer memory on decision-making accuracy.

A new benchmark called KidVis has been introduced to evaluate the visual perceptual capabilities of Multimodal Large Language Models (MLLMs), specifically assessing their performance against that of 6-7 year old children across six atomic visual capabilities. The results reveal a significant performance gap, with human children scoring an average of 95.32 compared to GPT-5's score of 67.33.

The introduction of the Rotation-Equivariant Anchor Transformer (REVNET) aims to enhance point cloud completion by addressing the limitations of existing methods that struggle with arbitrary rotations. This novel framework utilizes Vector Neuron networks to predict missing data in point clouds, which is crucial for applications relying on accurate 3D representations.

A new method called PRISM has been introduced to optimize the selection of training data for Multimodal Large Language Models (MLLMs), addressing the redundancy in rapidly growing datasets that increases computational costs. This self-pruning intrinsic selection method aims to enhance efficiency without the need for extensive training or proxy-based inference techniques.