Video-RAG: Visually-aligned Retrieval-Augmented Long Video Comprehension

FOCUS, a new keyframe selection module, has been introduced to enhance long video understanding by selecting query-relevant frames while adhering to strict token budgets. This model-agnostic approach formulates keyframe selection as a combinatorial pure-exploration problem, aiming to identify the most informative video segments without prior filtering.

A new framework called Perception Loop Reasoning (PLR) has been introduced to enhance video understanding by addressing the limitations of existing Video Reasoning LLMs, which often rely on a flawed single-step perception paradigm. This framework integrates a loop-based approach with an anti-hallucination reward system to improve the accuracy and reliability of video analysis.

Recent research has evaluated the performance of large vision language models (VLMs) in answering medical questions based on visual information, specifically using the EuropeMedQA Italian dataset. Four models were tested: Claude Sonnet 4.5, GPT-4o, GPT-5-mini, and Gemini 2.0 flash exp. The findings indicate varying degrees of visual grounding, with GPT-4o showing the most significant drop in accuracy when visual information was altered.

The introduction of LAST, or LeArning to Think in Space and Time, aims to enhance the capabilities of vision-language models (VLMs) by enabling them to better understand 3D spatial contexts and long video sequences using only 2D images as input. This approach contrasts with existing methods that typically address 3D and video tasks separately.

A novel framework called Visual Contrast Exploitation (VCE) has been proposed to enhance the safety of autoregressive image generation models, which have gained attention for their ability to create highly realistic images. This framework aims to address concerns regarding the generation of Not-Safe-For-Work (NSFW) content and copyright infringement by introducing a method for constructing contrastive image pairs that effectively decouple unsafe content from the generated images.

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.