An Image Is Worth Ten Thousand Words: Verbose-Text Induction Attacks on VLMs

A new evaluation framework for assessing the cultural interpretation capabilities of Vision-Language Models (VLMs) has been introduced, focusing on cross-cultural art critique. This tri-tier framework includes automated metrics, rubric-based scoring, and calibration against human ratings, revealing a 5.2% reduction in mean absolute error in cultural understanding assessments.

SwiftMem has been introduced as a query-aware agentic memory system designed to enhance the efficiency of large language model (LLM) agents by enabling sub-linear retrieval through specialized indexing techniques. This system addresses the limitations of existing memory frameworks that rely on exhaustive retrieval methods, which can lead to significant latency issues as memory storage expands.

PrivGemo has been introduced as a privacy-preserving framework designed for knowledge graph (KG)-grounded reasoning, addressing the risks associated with using private KGs in large language models (LLMs). This dual-tower architecture maintains local knowledge while allowing remote reasoning through an anonymized interface, effectively mitigating semantic and structural exposure.

A recent study has introduced Concept-Based Diversity (CBD), a highly efficient metric for image inputs that utilizes Vision-Language Models (VLMs) to enhance the performance of Deep Neural Networks (DNNs) through improved input selection. This approach addresses the computational intensity and scalability issues associated with traditional diversity-based selection methods.

A new offline reinforcement learning (RL) framework named STO-RL has been proposed to enhance policy learning from pre-collected datasets, particularly in long-horizon tasks with sparse rewards. By utilizing large language models (LLMs) to generate temporally ordered subgoal sequences, STO-RL aims to improve the efficiency of reward shaping and policy optimization.

Recent research highlights that while KV cache reuse can enhance efficiency in multi-agent large language model (LLM) systems, it can negatively impact the performance of LLM judges, leading to inconsistent selection behaviors despite stable end-task accuracy.

The introduction of VMMU, a Vietnamese Multitask Multimodal Understanding and Reasoning Benchmark, aims to assess the capabilities of vision-language models (VLMs) in interpreting and reasoning over visual and textual information in Vietnamese. This benchmark includes 2.5k multimodal questions across seven diverse tasks, emphasizing genuine multimodal integration rather than text-only cues.

A new approach called SubTTA has been proposed to enhance test-time adaptation (TTA) for Vision-Language Models (VLMs), addressing vulnerabilities to distribution shifts that can misguide adaptation through unreliable zero-shot predictions. SubTTA aligns the semantic subspaces of visual and textual modalities to improve the accuracy of predictions during adaptation.