LAST: LeArning to Think in Space and Time for Generalist Vision-Language Models

MASTEST, a multi-agent system utilizing large language models (LLMs), has been developed to enhance the testing of RESTful APIs, crucial for cloud-native application quality assurance. The system automates the entire API testing workflow, from generating test scenarios based on OpenAPI specifications to executing tests and analyzing responses for correctness and coverage.

OmniStruct has been introduced as a comprehensive benchmark to evaluate the capabilities of Large Language Models (LLMs) in generating structured outputs across various tasks, including information extraction, table generation, and function calling. This initiative aims to address the uncertainty regarding LLMs' performance in text-to-structure tasks, which are essential for diverse applications.

Large Language Models (LLMs) like GPT-4o have been evaluated for their effectiveness in assessing the difficulty of programming tasks, specifically through a comparison with a Light-GBM ensemble model. The study revealed that Light-GBM achieved 86% accuracy in classifying LeetCode problems, while GPT-4o only reached 37.75%, indicating significant limitations in LLMs for structured assessments.

A new architecture called Structured Cognitive Loop (SCL) has been introduced to address fundamental issues in large language model agents, such as entangled reasoning and memory volatility. SCL separates cognition into five distinct phases: Retrieval, Cognition, Control, Action, and Memory, while employing Soft Symbolic Control to enhance explainability and controllability. Empirical tests show SCL achieves zero policy violations and maintains decision traceability.

Recent research has focused on the latent reasoning capabilities of large language models (LLMs), specifically through a study on two-hop question answering. The investigation revealed that LLMs, including Llama 3 and GPT-4o, struggle with this basic reasoning task without employing chain-of-thought (CoT) techniques, which are essential for complex agentic tasks.

A new approach called TRIM has been introduced to address the high inference costs associated with Large Language Models (LLMs). This method optimizes language generation by allowing LLMs to omit semantically irrelevant words during inference, followed by reconstruction of the output using a smaller, cost-effective model. Experimental results indicate an average token saving of 19.4% for GPT-4o with minimal impact on evaluation metrics.

A recent study evaluated the safety of four leading multimodal large language models (MLLMs) under adversarial conditions, revealing significant differences in their vulnerability to harmful prompts. The models tested included GPT-4o, Claude Sonnet 3.5, Pixtral 12B, and Qwen VL Plus, with Pixtral 12B showing a harmful response rate of approximately 62%, while Claude Sonnet 3.5 demonstrated the highest resistance at around 10%.

Recent advancements in multimodal large language models (MLLMs) and video agent systems have led to the development of SciEducator, an innovative multi-agent system designed for scientific video comprehension and education. This system utilizes the Deming Cycle's iterative approach to enhance the understanding of complex scientific processes through tailored multimodal educational content.