Exploiting Synergistic Cognitive Biases to Bypass Safety in LLMs

Recent advancements in vision-language models (VLMs) have utilized large language models (LLMs) to achieve performance comparable to proprietary systems like GPT-4V. However, deploying these models on resource-constrained devices poses challenges due to high computational requirements. To address this, a new framework called Generation after Recalibration (GenRecal) has been introduced, which distills knowledge from large VLMs into smaller, more efficient models by aligning feature representations across diverse architectures.

Large language models (LLMs) are increasingly utilized in finance and economics, where their ability to understand chronology is critical. A study tested this capability through various chronological ordering tasks, revealing that while models like GPT-4.1 and GPT-5 can maintain local order, they struggle with creating a consistent global timeline. The findings indicate a significant drop in exact match rates as task complexity increases, particularly in conditional sorting tasks, highlighting inherent limitations in LLMs' chronological reasoning.

MedBench v4 is a new benchmarking infrastructure designed to evaluate Chinese medical language models, multimodal models, and intelligent agents. It features over 700,000 expert-curated tasks across various specialties, with evaluations conducted by clinicians from more than 500 institutions. The study assessed 15 advanced models, revealing that base LLMs scored an average of 54.1/100, while safety and ethics ratings were notably low at 18.4/100. Multimodal models performed even worse, indicating a need for improved evaluation frameworks in medical AI.

The research paper explores the challenge of false information and the effectiveness of large language models (LLMs) in verifying factual claims in English and Telugu. It presents a bilingual dataset and evaluates various approaches for classifying the veracity of claims. The study aims to enhance the efficiency of fact-checking processes, which are often labor-intensive and time-consuming.

Self-Examining Reinforcement Learning (SERL) is a proposed framework that addresses challenges in applying Reinforcement Learning (RL) to open-domain tasks. Traditional methods face issues with subjectivity and reliance on external rewards. SERL innovatively positions large language models (LLMs) as both Actor and Judge, utilizing internal reward mechanisms. It employs Copeland-style pairwise comparisons to enhance the Actor's capabilities and introduces a self-consistency reward to improve the Judge's reliability, aiming to advance RL applications in open domains.

10Cache is a new tensor caching and migration system designed to enhance the training of large language models (LLMs) in cloud environments. It addresses the challenges of memory bottlenecks associated with GPUs by optimizing memory usage across GPU, CPU, and NVMe tiers. By profiling tensor execution order and constructing prefetch policies, 10Cache improves memory efficiency and reduces training time and costs, making large-scale LLM training more feasible.

The integration of Large Language Models (LLMs) with 3D vision is revolutionizing robotic perception and autonomy. This approach enhances robotic sensing technologies, allowing machines to understand and interact with complex environments using natural language and spatial awareness. The review discusses the foundational principles of LLMs and 3D data, examines critical 3D sensing technologies, and highlights advancements in scene understanding, text-to-3D generation, and embodied agents, while addressing the challenges faced in this evolving field.

The article presents a new framework called GMAT, which enhances Multiple Instance Learning (MIL) for whole slide image (WSI) classification. By integrating vision-language models (VLMs), GMAT aims to improve the generation of clinical descriptions that are more expressive and medically specific. This addresses limitations in existing methods that rely on large language models (LLMs) for generating descriptions, which often lack domain grounding and detailed medical specificity, thus improving alignment with visual features.