Mixture of Attention Spans: Optimizing LLM Inference Efficiency with Heterogeneous Sliding-Window Lengths

HunyuanOCR has been introduced as a new open-source Vision-Language Model (VLM) designed for Optical Character Recognition (OCR) tasks, showcasing a lightweight architecture with 1 billion parameters. It has demonstrated superior performance in various OCR-related tasks, outperforming existing commercial APIs and larger models, and has secured first place in the ICDAR 2025 DIMT Challenge.

A recent study has introduced an automated pipeline for generating large-scale, psychologically-grounded multi-turn jailbreak datasets for Large Language Models (LLMs). This approach leverages psychological principles like Foot-in-the-Door (FITD) to create a benchmark of 1,500 scenarios, revealing significant vulnerabilities in models, particularly those in the GPT family, when subjected to multi-turn conversational attacks.

A new framework called Mosaic Pruning (MoP) has been introduced to enhance the generalizability of Sparse Mixture-of-Experts (SMoE) models, addressing the limitations of existing pruning methods that often lead to performance degradation across different domains. MoP employs a structured 'cluster-then-select' process to create a comprehensive set of experts, significantly reducing the static memory overhead associated with loading all experts during inference.

Recent research has highlighted significant vulnerabilities in Large Language Models (LLMs), particularly concerning prompt injection and jailbreaking attacks. This review categorizes various attack methods and evaluates defense strategies, including prompt filtering and self-regulation, to mitigate these risks.

The introduction of HERMES, a Heterogeneous Multi-stage LLM inference Execution Simulator, marks a significant advancement in optimizing inference pipelines for Large Language Models (LLMs). This tool addresses the limitations of existing simulators by accurately modeling diverse request stages, including Retrieval Augmented Generation (RAG) and key-value cache retrieval, across complex hardware architectures.

A new framework called PersonaPulse has been introduced to optimize prompts for Large Language Models (LLMs), enhancing their ability to express realistic personality traits. This approach iteratively refines role-play prompts while using a situational response benchmark for evaluation, demonstrating improved performance over previous methods based on psychological personality descriptions.

A systematic study has been conducted on compression ordering for large language models (LLMs), specifically focusing on the Qwen2.5 3B model. The research evaluates various compression techniques such as knowledge distillation, structured pruning, and low-bit quantization, analyzing their performance both independently and in combination. The findings indicate that quantization offers the highest standalone compression, while the sequence of techniques significantly impacts the final model quality.

A recent study published on arXiv investigates prompt fairness in Large Language Models (LLMs), revealing significant disparities in response quality based on how prompts are phrased by different users. The research employs information-theoretic metrics to assess subgroup sensitivity and cross-group consistency, highlighting structural inequities in model behavior across various demographic subgroups.